# Ten steps for agent-made video, with proof Built 2026-11-03 from 7,605 X posts (5,487 from Oct 27 – Nov 3, 2026) and 1,470 varg how-to cards. Every rule cites the posts or cards it came from; numbers are what X or the card reported on the build date. ## 01. References Find real film, design and pop-culture frames for the look you want. Agents average everything out; references pull them off the middle. ### 01.1 Before building, find 1-2 real references in the target style and tell the agent to match them instead of describing the style. Without a reference the agent falls back to its default look (centered text, gradient background, everything fading in). A reference gives it pacing, type and transitions to copy; naming a style works better than describing one. Steal taste from work that already worked. Do: - Pull film stills from ShotDeck or Frameset, or search FrameThrower by lighting, lens, shot size, color and mood; boards from Savee or Cosmos; transitions from Eyecannndy. - For launch films, search Refs over MCP for pieces measured shot by shot and pull their timing and plan. - For design, pick 3 to 5 sites on Godly, Refero or One Page Love, drop the screenshots in and have the agent say what makes each look good. - Pick 1-2 references and tell the agent to match them; attach screenshots as a quality bar and match the polish, not the brand. - Keep references as deep links to shot.cafe, film-grab and eyecannndy, never as copied frames, and trace every one to its source. - Turn the parts you picked into a DESIGN.md and import it into CLAUDE.md. Done when: The render has been scored against the reference film (cut timing, brightness, motion, loudness, format) and holds up next to the references. ```text style references: [links or a folder of videos whose motion i like] ``` Proof: @rexan_wong (24,918 followers), 18,188 saves / 6,888 likes / 647,682 views, 2026-09-26 https://x.com/rexan_wong/status/2103707054108299437; @EXM7777 (156,736 followers), 562 saves / 294 likes / 41,310 views, 2026-08-10 https://x.com/EXM7777/status/2086834484390019153; varg how to skill "FrameThrower Film References", 1 stars https://varg.ai/how-to/skills/framethrower; varg how to skill "Video references by Refs" https://varg.ai/how-to/skills/refs-video-references; varg how to how_to "How to make a launch video from a reference film" https://varg.ai/how-to/make-a-launch-video-from-a-reference; @Voxyz_ai (27,685 followers), 1,400 saves / 781 likes / 104,389 views, 2026-10-03 https://x.com/Voxyz_ai/status/2106474370860548341; @0xMovez (47,109 followers), 1,785 saves / 1,283 likes / 163,211 views, 2026-10-29 https://x.com/0xMovez/status/2115870377523949687; varg how to skill "Reference Finder" https://varg.ai/how-to/skills/craft-reference-finder; @ethancole_ai (119,804 followers), 266 saves / 275 likes / 146,976 views, 2026-11-02 https://x.com/ethancole_ai/status/2117283871272092007; varg how to skill "Check a render against its reference by Refs" https://varg.ai/how-to/skills/refs-check-render; @everestchris6 (27,118 followers), 2,836 saves / 1,062 likes / 128,536 views, 2026-09-29 https://x.com/everestchris6/status/2104965415164399895 ### 01.2 Break each reference into a written spec (compositions, type, scene changes, pacing, motion, audio, lens, light, palette, grain) before the agent builds. Sources inspect the reference's compositions, typography, scene changes, motion and audio before building, and pull structure, pacing and timing so you are not guessing what made it work. A vision model names lens, light, palette and grain in one locked paragraph. Keep the structure, not the footage. Do: - Download the reference by link (Instagram, TikTok, X, YouTube) or pull it with yt-dlp. - Tile frames from the reference video into one image and pull out its pacing, type and transitions as code. - Have a vision model name the lens, light, palette and grain of the picks in one locked paragraph pasted into every prompt. - Inspect compositions, typography, scene changes, motion and audio; ask the user only for what cannot be recovered from the reference. - Keep the structure, not the original footage. - Play the original and the rebuild side by side in sync and scrub to see where the timing drifts. Done when: A side-by-side, in-sync comparison of original and rebuild has been scrubbed for timing drift. ```text Use the attached PHYLLA video as the primary reference. Inspect its compositions, typography, scene changes, motion, and audio before building. Ask me for any available original image assets, PHYLLA logo, brand fonts, Higgsfield/Katana branding, and soundtrack. Ask only for information that cannot be recovered from the reference. ``` Proof: @Voxyz_ai (27,685 followers), 1,440 saves / 711 likes / 130,399 views, 2026-09-27 https://x.com/Voxyz_ai/status/2104194355556671987; @EXM7777 (156,736 followers), 562 saves / 294 likes / 41,310 views, 2026-08-10 https://x.com/EXM7777/status/2086834484390019153; varg how to prompt "30 s AI-model launch film rebuilt from a reference in code" https://varg.ai/how-to/prompts/x-2108370085178138694; @YourAlphaMom (5,473 followers), 1,227 saves / 2,098 likes / 637,589 views, 2026-09-11 https://x.com/YourAlphaMom/status/2098541423062548561; varg how to how_to "How to remake a reference video with AI" https://varg.ai/how-to/remake-a-reference-video; @flxrnc (3,305 followers), 2,763 saves / 1,330 likes / 106,917 views, 2026-11-02 https://x.com/flxrnc/status/2117318336837017846; @Asteri_eth (3,340 followers), 370 saves / 257 likes / 13,101 views, 2026-10-28 https://x.com/Asteri_eth/status/2115496727893410053; @Mho_23 (34,382 followers), 1,517 saves / 1,075 likes / 100,836 views, 2026-07-29 https://x.com/Mho_23/status/2082600596457341398 ### 01.3 Give every reference one labelled role and state what it must keep, what it must leave out, and what it does not define. Each reference gets a role ('match exactly' vs 'reference only for') and a keep/leave-out list. A face reference defines who the person is, not pose or lighting. A raw dance video hands the model ten variables when you wanted one; a depth map leaves only the motion. Do: - Tag each image by role: identity, face, hair and wardrobe; location layout and lighting; 'look only', 'set only', 'prop only'. - Write that the face reference defines WHO the person is, not the pose, location, expression, clothing, lighting or composition. - Refer to a video only for pacing, light or camera movement; do not copy its people, places or lines. - Crop a dense character sheet into 2-3 clean panels (hero portrait, front full-body) as separate references. - For the grade, feed a ShotDeck shot card as a reference image and write 'scene tone strictly per image N'. - For motion transfer, feed a depth map of the source footage as the motion reference; never reproduce a sheet's side-by-side layout. Done when: Every reference in the prompt has a stated role and keep/leave-out list, and the output shows no copied sheet layout, panel borders or gray backdrop. ```text Image 1 is the lead character: keep her face, hair and outfit in every shot. Image 2 is the hero product ({PRODUCT}, e.g. a sofa): keep its shape, colour and fabric exactly. Image 3 is the store front: keep its layout, no readable text except {BRAND}. Refer to Video 1 only for pacing and warm cinema light; do not copy its people, places or lines. ``` ```text Use the uploaded reference image as the identity reference for the person. Preserve the same recognizable facial identity, facial structure, skin tone, hair characteristics, age, and overall appearance throughout all nine images. The reference image defines WHO the person is, not the pose, location, expression, clothing, lighting, or composition. Create ONE single image formatted as a clean 3×3 grid containing exactly 9 separate smartphone photographs of the same person. ``` ```text scene tone strictly per image 3 ``` Proof: varg how to prompt "21:9 fisheye key frame with four role-tagged references" https://varg.ai/how-to/prompts/yt-kling4-gpt-image-crash-keyframe; @iamsofiaijaz (15,879 followers), 76 saves / 136 likes / 6,157 views, 2026-10-28 https://x.com/iamsofiaijaz/status/2115406342802506033; varg how to prompt "Animal-knight fantasy short with dialogue, 21:9" https://varg.ai/how-to/prompts/x-2116728095323677043; @swiat_ai (39,080 followers), 1,521 saves / 636 likes / 553,382 views, 2026-09-17 https://x.com/swiat_ai/status/2100482637341012201; varg how to prompt "Furniture chain's first all-AI cinema film" https://varg.ai/how-to/prompts/yt-4r8mcwjrhsm; varg how to prompt "Heist-movie reply ad to a copycat restaurant" https://varg.ai/how-to/prompts/x-2083364450900549639; @meAsifAi (4,494 followers), 3,450 saves / 1,834 likes / 104,416 views, 2026-09-10 https://x.com/meAsifAi/status/2098429368053960936; @Iancu_ai (3,183 followers), 383 saves / 274 likes / 11,951 views, 2026-07-22 https://x.com/Iancu_ai/status/2079936184516497702; @NexlowX (1,536 followers), 868 saves / 691 likes / 120,534 views, 2026-08-31 https://x.com/NexlowX/status/2094494202587398616; @Mavern_0x (3,413 followers), 14 saves / 48 likes / 2,662 views, 2026-11-02 https://x.com/Mavern_0x/status/2117403866194903400; @bmx_ai13 (10,642 followers), 772 saves / 523 likes / 80,827 views, 2026-08-09 https://x.com/bmx_ai13/status/2086544422389547488; varg how to skill "Keep One Character Consistent Across Seedance 2.5 Action Shots" https://varg.ai/how-to/skills/crow-line-seedance-action-film Top 10: 1. Pick 1-2 reference videos in the style you want and tell the agent to match them; naming a style works better than describing it. (https://x.com/rexan_wong/status/2103707054108299437) 2. Render one still beside the reference picture and fix the differences round after round before making any video. (https://x.com/elvar_official/status/2116192984299344372) 3. Keep the full character sheet as the design lock but feed the video model 2-3 cropped panels (hero portrait, front full-body) as separate references; small faces drift. (https://x.com/meAsifAi/status/2098429368053960936) 4. State what a face reference does not define: it fixes who the person is, not pose, location, expression, clothing, lighting or composition. (https://x.com/swiat_ai/status/2100482637341012201) 5. Have a breakdown subagent tile frames from a reference video into one image and write up its pacing, type and transitions as code. (https://x.com/Voxyz_ai/status/2104194355556671987) 6. Pick 3-5 sites you like on Godly, Refero or One Page Love, drop their screenshots into Claude Code and have it explain what makes each look good. (https://x.com/Voxyz_ai/status/2106474370860548341) 7. Convert source dance footage to a depth map and feed it as the motion reference, with the character sheet as identity reference, so only motion transfers. (https://x.com/NexlowX/status/2094494202587398616) 8. Have a vision model name the lens, light, palette and grain of your chosen references in one locked paragraph, and paste it into every prompt. (https://x.com/EXM7777/status/2086834484390019153) 9. Steal taste from work that already worked: film stills from Frameset and ShotDeck, boards from Savee and Cosmos, transitions from Eyecannndy. (https://x.com/EXM7777/status/2086834484390019153) 10. Give each reference a role ('match exactly' or 'reference only for') and list what to keep or leave out, then add a scene paragraph, a grade and an Avoid line. (https://varg.ai/how-to/prompts/yt-kling4-gpt-image-crash-keyframe) ## 02. Moodboard Put 10–20 of them in one folder and have Claude write down what they share: palette, light, framing, type. ### 02.1 Put references in one folder, have the agent write what they share as a style bible, and import it so later outputs follow it. Without a reference the agent falls back to its default look: centered text, gradient background, everything fading in. A reference gives it pacing, type and transitions to copy; what the best references share became rules; a DESIGN.md imported into CLAUDE.md makes every page after it follow it. Do: - Start a folder of template videos and screenshots from creators whose work you like, collecting references before generating anything; save videos by link or Pinterest images locally with media-download. - Tile frames from each reference video into one image, pull out its pacing, type and transitions, and have the agent say what makes each reference look good. - Extract each reference's 6 dominant colours with k-means and record them as hex codes with percentages. - Ask: "extract the art direction: mood, motion, typography, pacing. write it down as a style bible", and record WHY each animation exists. - Turn what the best references share into rules, add house rules (no template layouts, no stock gradients), save it as DESIGN.md and import it into CLAUDE.md. - After each decent result, update the style skill with what worked. Done when: A style bible / DESIGN.md exists and is imported into CLAUDE.md, and a second agent with fresh context grades the output as holding up next to the references. ```text extract the art direction: mood, motion, typography, pacing. write it down as a style bible ``` ```text reverse-engineer the art direction: mood, typography, pacing, and WHY each animation exists. save it as a style bible ``` Proof: @0xMiraqle (4,886 followers), 1,684 saves / 763 likes / 138,757 views, 2026-07-12 https://x.com/0xMiraqle/status/2076217897005855093; @0xMiraqle (4,886 followers), 1,581 saves / 755 likes / 207,511 views, 2026-07-15 https://x.com/0xMiraqle/status/2077199463278395732; @rexan_wong (24,918 followers), 18,188 saves / 6,888 likes / 647,682 views, 2026-09-26 https://x.com/rexan_wong/status/2103707054108299437; @sinanyorulmazz (861 followers), 612 saves / 370 likes / 40,084 views, 2026-10-31 https://x.com/sinanyorulmazz/status/2116531845839851723; @Voxyz_ai (27,685 followers), 1,400 saves / 781 likes / 104,389 views, 2026-10-03 https://x.com/Voxyz_ai/status/2106474370860548341; @Voxyz_ai (27,685 followers), 1,440 saves / 711 likes / 130,399 views, 2026-09-27 https://x.com/Voxyz_ai/status/2104194355556671987; @ashebytes (45,332 followers), 973 saves / 591 likes / 44,357 views, 2026-10-27 https://x.com/ashebytes/status/2115215620279263329; @everestchris6 (27,118 followers), 2,836 saves / 1,062 likes / 128,536 views, 2026-09-29 https://x.com/everestchris6/status/2104965415164399895; varg how to skill "Local-first video downloader: private, no subscription" https://varg.ai/how-to/skills/media-download; @aimikoda (34,313 followers), 337 saves / 402 likes / 22,088 views, 2026-10-20 https://x.com/aimikoda/status/2112590477144797190; @poistudioltd (3,245 followers), 489 saves / 398 likes / 15,412 views, 2026-08-08 https://x.com/poistudioltd/status/2086013571218149595 ### 02.2 Write every style dimension as something checkable: hex palette with ground texture, numeric light, exact type spec, concrete character words, named recording format. A brief in numbers gives the agent something to check against. One shared format sets the palette as hex on grained paper and captions by font, weight, size and position; character gets two or three concrete words, not clean, modern or premium; period looks name the format, not a mood. Do: - List the palette as exact hex values plus the ground texture, e.g. warm cream paper #EBE7E1 with fine noise grain, ink #1E1C1A, red #D7372C, grey #8A857E. - Write light in numbers: time of day and sun angle, colour temperature in Kelvin, focal length; motivate each light with a window, lamp, monitor or streetlight. - Fix caption type as font, weight, size and position, e.g. Instrument Sans 500 at 62px at x 120, y 196. - Describe character in two or three concrete words (not clean, modern, sleek, minimal or premium) and give rich colour to only one or two focus spots. - Limit the palette: 2–4 main colours from neighbouring hues and warm/cold contrast, plus one high-saturation accent point. - For a period or device look, name the recording format and its artifacts, not a mood; add halation and film grain with Dehancer. Done when: Every palette, light, type and texture line in the style file is a number, hex code, named light source or named format the agent can check a frame against. ```text A 1080x1920 explainer reel on warm cream paper (#EBE7E1 with fine noise grain). Colours: ink #1E1C1A, red #D7372C, grey #8A857E. Cards are #FBFAF8 with a 3px ink border, 20px corners and a halftone shadow (1.7px ink dots on an 8px grid, offset 13px left and 12px down). Captions sit top left (x 120, y 196, 850px wide) in Instrument Sans 500 at 62px. ``` Proof: @finnardley (3,467 followers), 1,808 saves / 986 likes / 58,721 views, 2026-10-30 https://x.com/finnardley/status/2116010421740007894; varg how to prompt "9:16 explainer reel with a talking cartoon presenter" https://varg.ai/how-to/prompts/x-2116010421740007894; @QibazX (1,076 followers), 111 saves / 77 likes / 6,736 views, 2026-10-28 https://x.com/QibazX/status/2115490139631743316; @Voxyz_ai (27,685 followers), 241 saves / 118 likes / 10,068 views, 2026-10-31 https://x.com/Voxyz_ai/status/2116515537496993856; @beginnersblog1 (29,874 followers), 167 saves / 107 likes / 4,588 views, 2026-11-02 https://x.com/beginnersblog1/status/2117355436789977462; varg how to skill "FLUX 3 Archival Formats", 101 installs, 135 stars https://varg.ai/how-to/skills/flux-3-archival-formats; @Voxyz_ai (27,685 followers), 6,914 saves / 2,903 likes / 645,020 views, 2026-09-26 https://x.com/Voxyz_ai/status/2103977414711767244; @xiaoxiaodong (30,592 followers), 472 saves / 424 likes / 126,810 views, 2026-08-19 https://x.com/xiaoxiaodong/status/2090016985883213883; @godswayfoundinc (3,213 followers), 1,369 saves / 928 likes / 89,543 views, 2026-08-27 https://x.com/godswayfoundinc/status/2092819297781334343 ### 02.3 Lock the look across shots: paste the same style blocks verbatim into every clip and give every reference one stated job. Clips share no memory, so locked blocks pasted into every clip carry the style. A reference image defines who, not pose, lighting or composition; a reference video serves only camera, grade and pacing; a brand guide page only colours, type and logo. Shots invented without stills look more AI. Do: - Keep the style as locked blocks and paste them verbatim into every single clip prompt (a 'Master (include with every clip)' section). - Write a role for each reference image: what it defines and what not to copy (backdrop, pose, flat studio lighting); the moodboard sets drawing treatment only, not characters, costumes or story. - Pass the brand guide page as an image reference used for colours, type and logo only; refer to a reference video only for camera movement, grade and pacing. - Set the look with stills before video and approve the look on cheap stills first. - Grade every shot to one reference with color-matcher. - Make one colour reference frame per scene, have a continuity agent check them together, then play every clip in a review scene with fixed camera and the concept art pinned in the corner. Done when: Every clip prompt carries the identical locked style block, every reference has a stated role, and every clip has been checked in the review scene against the pinned concept art. ```text locked blocks, paste verbatim into every single clip STYLE: "Real documentary UGC footage, static locked camera on a dry-erase whiteboard, one real human hand and forearm only, never show a face or body. A single black marker is used for the entire clip, no color changes, no marker switching. Marker ink renders as genuine dry-erase marker, not a digital text or motion-graphic overlay. Handwriting is plain block capitals, legible at a glance. Icons stay minimal and schematic, one simple recognizable shape only. No captions, no subtitles, no on-screen text beyond what the hand physically writes ``` ```text Use the uploaded reference image as the identity reference for the person. Preserve the same recognizable facial identity, facial structure, skin tone, hair characteristics, age, and overall appearance throughout all nine images. The reference image defines WHO the person is, not the pose, location, expression, clothing, lighting, or composition. ``` ```text Refer to Video 1 only for camera movement, grade and pacing; do not copy its people, places or lines. ``` Proof: varg how to prompt "Whiteboard explainer ad, 9 clips chained by last frame" https://varg.ai/how-to/prompts/x-2086118407708250149; @sandmrk (10,218 followers), 27 saves / 43 likes / 3,665 views, 2026-11-02 https://x.com/sandmrk/status/2117243590661562547; @swiat_ai (39,080 followers), 1,521 saves / 636 likes / 553,382 views, 2026-09-17 https://x.com/swiat_ai/status/2100482637341012201; @AiwithBloodline (7,256 followers), 80 saves / 128 likes / 5,892 views, 2026-10-29 https://x.com/AiwithBloodline/status/2115841225106604469; varg how to prompt "Fictional energy-drink brand built from one brand skill" https://varg.ai/how-to/prompts/x-2095902585194475945; varg how to prompt "The RealReal's AI man who can't prove he is real" https://varg.ai/how-to/prompts/yt-m0pudxvau5c; varg how to prompt "30s scooter run through NYC traffic, no slow motion" https://varg.ai/how-to/prompts/yt-kling4-scooter-traffic-montage; varg how to how_to "How to make video ad variants for testing with AI" https://varg.ai/how-to/make-video-ad-variants-for-testing; @0xwhrrari (23,306 followers), 2,635 saves / 1,406 likes / 93,462 views, 2026-11-02 https://x.com/0xwhrrari/status/2117240315069276186; @rewind02 (7,898 followers), 782 saves / 611 likes / 27,451 views, 2026-10-29 https://x.com/rewind02/status/2115779810932728259; @IamEmily2050 (49,235 followers), 191 saves / 210 likes / 9,353 views, 2026-09-24 https://x.com/IamEmily2050/status/2103173656356749367; @nrqa__ (76,449 followers), 64 saves / 200 likes / 71,522 views, 2026-10-22 https://x.com/nrqa__/status/2113304737236070789 Top 10: 1. Give the agent 1–2 reference videos and tell it to match them; naming a style works better than describing one. (https://x.com/rexan_wong/status/2103707054108299437) 2. Scope each reference image to what it defines and tell the model not to take pose, location, lighting or composition from it. (https://x.com/swiat_ai/status/2100482637341012201) 3. Write the palette as exact hex values plus the ground texture, e.g. cream paper with fine noise grain and three ink colours. (https://x.com/finnardley/status/2116010421740007894) 4. Feed 2–3 loved references with one line asking the agent to extract mood, motion, typography and pacing and write it down as a style bible. (https://x.com/0xMiraqle/status/2076217897005855093) 5. Turn the traits you pick into a DESIGN.md and import it into CLAUDE.md so every later output follows it. (https://x.com/Voxyz_ai/status/2106474370860548341) 6. Start a folder of template videos from creators whose reels or long-form snippets you like, and keep adding to it. (https://x.com/ashebytes/status/2115215620279263329) 7. Build a review scene (neutral floor, fixed camera, concept art pinned in the corner) and check every clip in it. (https://x.com/rewind02/status/2115779810932728259) 8. Compare a set of high-rated references and turn what the best ones share into written rules. (https://x.com/sinanyorulmazz/status/2116531845839851723) 9. Keep the style as locked blocks and paste them verbatim into every clip prompt. (https://varg.ai/how-to/prompts/x-2086118407708250149) 10. Get a period or device look by naming the recording format and its artifacts, not a mood. (https://varg.ai/how-to/skills/flux-3-archival-formats) ## 03. Script Story beats, timings, one recurring motif, written before anything moves. ### 03.1 Write story.md before anything moves: one point, one hero number, three beats and one motif, with each step saved as its own file. Most AI videos look the same because people skip straight to 'make me a video'. Small jobs that each hand one file to the next are easy to check, and nothing gets re-rolled. One recurring motif is why a code-made history film feels designed rather than generated. Do: - Watch a reference video the user likes and write spec.md: why it works, its pacing, where the first motion lands and the cut rhythm. - Pull the real numbers into data.json and keep the source next to every figure, so nothing on screen is made up. - Write story.md: one point, one hero number, three beats (Stageshow plans the whole piece in 3-5 beats) before any motion code. - In story.md, pick ONE motif (a glowing word, a dot, a line) that carries the whole film and let it evolve; end on a closing line that echoes the opening. - Keep timings in timing.ts so a plain note like 'title is a beat late' changes one line and nothing else breaks. - Render stills, check them against spec.md and data.json, and wait for the user's approval before shipping final.mp4. Done when: story.md holds one point, one hero number, three beats and one named motif, and the rendered stills have been checked against spec.md and data.json and approved. ```text one point, one hero number, three beats. before anything moves ``` ```text Step 1: write a 7-part script of about 380 words with one closing line that echoes the opening. Show it to me before building. Step 2: voice it with an ElevenLabs narrator ({VOICE_STYLE}, calm, mid pace); get word timestamps; the voice drives every cut. Step 3: pick ONE motif that carries the whole film ({MOTIF}, e.g. a glowing word, a dot, a line) and let it evolve in ``` Proof: @francescoinweb3 (1,741 followers), 3,770 saves / 2,072 likes / 162,093 views, 2026-10-17 https://x.com/francescoinweb3/status/2111393917899808880; @francescoinweb3 (1,741 followers), 1,234 saves / 706 likes / 55,949 views, 2026-10-16 https://x.com/francescoinweb3/status/2111190424647782538; @shannholmberg (37,522 followers), 6,386 saves / 2,686 likes / 247,561 views, 2026-11-01 https://x.com/shannholmberg/status/2116957574008590445; @ScottyBeamIO (78,454 followers), 289 saves / 219 likes / 18,277 views, 2026-11-01 https://x.com/ScottyBeamIO/status/2116919908474789939; @Y0urDumble (783 followers), 2,396 saves / 1,249 likes / 79,694 views, 2026-10-27 https://x.com/Y0urDumble/status/2115193601458151552; varg how to prompt "3-minute code-only history film made in Claude Code" https://varg.ai/how-to/prompts/x-2102844654169575547 ### 03.2 Hook with the most painful consequence, keep one real lesson, cut all filler, and withhold the answer until the last few seconds. MethodByVid rips transcripts of an expert's shorts, keeps the actual lesson, hooks with the most expensive or painful consequence and withholds the final answer until the end; a weakest-to-strongest ranking gives viewers a reason to stay. Withholding information for its own sake, without a reason, reads as a flaw. Do: - Rip transcripts from an expert's shorts or long-form videos and reformat the advice into actionable points. - Write the script so it keeps the actual lesson and cuts every ounce of filler. - Open on the most expensive or painful consequence and hold the final answer until the last few seconds, with a story reason for holding it back. - For list formats, rank items from weakest to strongest so viewers stay until the end and argue in the comments. - Size the voiceover to the runtime at normal conversational pace, around 120 to 150 words per minute. - Read the draft and change everything that sounds like AI slop. Done when: The opening line states the costliest consequence, the answer arrives only in the last few seconds, the word count fits the runtime at 120-150 words per minute, and a read-through finds no line that sounds like AI slop. ```text tell chatgpt to turn that actionable advice into a script that keeps the actual financial lesson, cuts every ounce of filler, hooks with the most expensive/painful consequence, and withholds the final answer until the last few seconds ``` Proof: @MethodByVid (3,078 followers), 6,276 saves / 3,066 likes / 510,994 views, 2026-09-24 https://x.com/MethodByVid/status/2103175381104935019; @MethodByVid (3,078 followers), 6,404 saves / 3,926 likes / 334,702 views, 2026-09-20 https://x.com/MethodByVid/status/2101760054903087141; @crowdeacon (1,627 followers), 77 saves / 318 likes / 40,790 views, 2026-10-24 https://x.com/crowdeacon/status/2113977109437939935; @brivael (240,056 followers), 187 saves / 2,172 likes / 42,444 views, 2026-10-28 https://x.com/brivael/status/2115553835443732554 ### 03.3 Turn the script into numbered, timestamped storyboard panels, each with VISUAL, ACTION and DIALOGUE and the verbatim voiceover line. Ciri_ai's 12-panel sheet is a fill-in-the-blank storyboard a video model can follow shot by shot (2,774 bookmarks). OlatundeAI keeps the script exactly as written next to each shot so visuals match during editing, and repeats each character's description verbatim in every shot. Do: - Add header boxes for Duration, Style, Audience and Audio Suggestion. - Lay out numbered panels in a grid, each with a timestamp range such as 0-2.5s, 2.5-5s. - Under every panel write VISUAL, ACTION and DIALOGUE. - Preserve the script exactly as written and put the matching script/voiceover line next to each shot. - Identify recurring characters first and repeat each one's established description verbatim in every shot. - When handing the sheet to a video model, tell it to follow the sheet's shot order and camera notes only. Done when: Every panel has a number, a timestamp, VISUAL/ACTION/DIALOGUE text and its script line, and the panel timestamps run back to back across the stated duration. ```text Create a clean, premium storyboard infographic for a product ASMR unboxing video. Design specifications: White background with a modern, minimalist layout. Title at the top: STORYBOARD Product name in bold green (replace with: [PRODUCT NAME]) Subtitle: ASMR UNBOXING Include four information boxes below the title: Duration: 30 Seconds Style: POV Hands, ASMR Audience: 18–35 years old Audio Suggestion: Soft ASMR • Crinkles • Tapping • Pouring Create 12 storyboard panels arranged in a 3-column × 4-row grid. Each panel should contain: A numbered green badge (1–12) Timestamp (0–2.5s, 2.5–5s, etc.) A realistic cinematic image showing the scene Handwritten white doodle text like: TA-DA! POP! WOW! PICK! CRINKLE~ TEAR! POUR~ HOT WATER STIR~ SMOOTH! BEST! Three text sections below every image: VISUAL: ACTION: DIALOGUE: ``` Proof: @Ciri_ai (11,908 followers), 2,773 saves / 2,207 likes / 301,002 views, 2026-07-12 https://x.com/Ciri_ai/status/2076202198888022135; varg how to prompt "12-panel ASMR unboxing storyboard sheet that becomes the ad" https://varg.ai/how-to/prompts/x-2076202198888022135; @OlatundeAI (6,183 followers), 45 saves / 34 likes / 3,192 views, 2026-10-30 https://x.com/OlatundeAI/status/2116043943892734000; varg how to prompt "Coconut-water spot made from one storyboard sheet" https://varg.ai/how-to/prompts/x-2094760813794111843; varg how to skill "Storyboard Creation", 11 installs, 781 stars https://varg.ai/how-to/skills/inference-sh-storyboard-creation; varg how to skill "Turn a Script Into a Shot List and AI Video Prompts" https://varg.ai/how-to/skills/cinematic-director-shot-list Top 10: 1. Rewrite source advice into a script that keeps the lesson, cuts all filler, hooks with the most painful consequence and withholds the answer until the end. (https://x.com/MethodByVid/status/2103175381104935019) 2. Write story.md before anything moves: one point, one hero number, three beats. (https://x.com/francescoinweb3/status/2111393917899808880) 3. Use bodily actions, such as sniffles between lines, as punctuation in dialogue. (https://x.com/second_bassoon/status/2117307653320007954) 4. Make a storyboard sheet of numbered panels, each with a timestamp and VISUAL, ACTION and DIALOGUE notes, plus duration, style, audience and audio boxes. (https://x.com/Ciri_ai/status/2076202198888022135) 5. Keep a facts file and treat it as the only source for any number, name or claim that appears on screen. (https://x.com/everestchris6/status/2104965415164399895) 6. Write the script and the shot plan, and show both to the user before spending anything. (https://x.com/finnardley/status/2117390487804068334) 7. Give each clip its own hook, suspense and release, stacking them while every clip stays true to the premise. (https://x.com/LinusEkenstam/status/2116850561832149323) 8. For a 30-60 s explainer, write a 2-line hook, 3-7 points and one closing line, and show the script before recording. (https://x.com/finnardley/status/2116010421740007894) 9. Lay out 3 storyboard options in one image for the user to pick, then make one still per scene for the chosen one. (https://x.com/Voxyz_ai/status/2104194355556671987) 10. Write dialogue with subtext: let characters talk around the point, e.g. 'you seem too young to be here' instead of 'how old are you?'. (https://x.com/diraism/status/2111414191940882724) ## 04. Sound and voice Lock the voice, dialogue and music first, and cut the picture to them. Half of what makes a video feel real is what you hear. ### 04.1 Generate and lock the voice before the picture: separate TTS, whole-second clips, exact script in prompt, and cut to its word timestamps. ZssBecker one-shots about 90% of talking ads by generating the voice separately and locking it; unlocked audio lets Seedance add or adjust speech to fit movement. Seedance also casts a voice from the character's look, so a screen test is where you hear and lock it. Do: - Generate every spoken line separately in ElevenLabs (or record it) before any video generation. - Give each character a screen test first (plain backdrop, slow push-in, short monologue to camera) to hear and lock the voice Seedance builds from their look. - Trim each reference audio file to whole seconds (6, 8, 10, 15) with no leftover tail. - Paste the exact text of that audio into the video prompt, lip-sync the character to it as @Audio1, and keep the audio locked over the whole clip. - For narrated pieces, get word timestamps from the voiceover and let the voice drive every cut. - Check the generated speech against the source audio and script: Seedance can still speed, slow or change it, and stray words break immersion. Done when: Each character's voice is locked from a screen test, the clip's speech matches the locked audio and script with no added, sped-up or slowed words, and cuts follow the voiceover's word timestamps. ```text A photoreal selfie video, 3:4, 22 mm, in the look of a low-quality impromptu phone selfie video: the man from @Image1 filming himself with the camera in his own hand while he walks through the kitchen in @Image2, talking to the viewer the whole way, lip-syncing to @Audio1, his facial expressions driven by @Video1 and kept small and natural. ``` ```text Step 2: voice it with an ElevenLabs narrator ({VOICE_STYLE}, calm, mid pace); get word timestamps; the voice drives every cut. ``` Proof: @ZssBecker (1,325,986 followers), 3,135 saves / 1,050 likes / 136,548 views, 2026-10-22 https://x.com/ZssBecker/status/2113320998527221875; @ZssBecker (1,325,986 followers), 3,135 saves / 1,050 likes / 136,548 views, 2026-10-22 https://x.com/ZssBecker/status/2113497202702115226; @Kiber_Alla (1,756 followers), 670 saves / 560 likes / 66,930 views, 2026-07-19 https://x.com/Kiber_Alla/status/2078814921756754352; varg how to prompt "3-minute code-only history film made in Claude Code" https://varg.ai/how-to/prompts/x-2102844654169575547; varg how to skill "ElevenLabs Text to Speech", 16,900 installs, 510 stars https://varg.ai/how-to/skills/elevenlabs-text-to-speech; @godswayfoundinc (3,213 followers), 76 saves / 76 likes / 3,954 views, 2026-10-20 https://x.com/godswayfoundinc/status/2112506617732153757; @MrOnsase (63,446 followers), 1 saves / 63 likes / 24,452 views, 2026-10-24 https://x.com/MrOnsase/status/2114045312805306555 ### 04.2 Measure the music, never trust ears: licensed track, librosa beat grid checked below 120 Hz, big moves on beats, one SFX per cut. The agent cannot hear, so it must measure; librosa sometimes locks onto the off-beat. A music-video pipeline starts from a librosa beat map of tempo, beats and onsets, and one creator sources a royalty-free ~120 BPM song with a drop from Mixkit, free for commercial use. Do: - Install ffmpeg and librosa; use the user's track or a Mixkit free-license track, and note the source and license. - Run librosa to pull tempo, beats and onsets into a beat map. - Plot the waveform below 120 Hz against the beat grid, zoom into a few bars, and shift the whole grid if the kicks don't line up. - Put scene cuts and title entrances on beats; if the animation can't change, move where the music starts. - Give every cut and big move one ready-made SFX file (whoosh, pop, click, typing), taking the project's sound library first. Done when: Kick drums line up with the beat grid in the low-frequency plot, and every scene cut and title entrance sits on a beat. ```text "Add sound to this video. You can't hear, so measure every step and don't go by feel. Install ffmpeg and librosa first if they're missing. 1. Use licensed music: a track I give you, or one under Mixkit's free license. Note the source and the license. Don't synthesize the music or the sound effects in code. 2. Find the beats with librosa. It sometimes locks onto the off-beat, so plot the waveform below 120 Hz against the beat grid and zoom in on a few bars. If the kick drums don't line up, shift the whole grid. 3. Put the big moves, like scene cuts and titles appearing, on the beat. If you can change the animation, change it. If you can't, move where the music starts. 4. Give every cut and big move one ready-made sound effect file: whoosh for movement, pop for things appearing, click for clicks, typing sounds for typing. ``` Proof: @Voxyz_ai (27,685 followers), 1,498 saves / 825 likes / 111,314 views, 2026-10-11 https://x.com/Voxyz_ai/status/2109358372873183564; @Asteri_eth (3,340 followers), 65 saves / 74 likes / 3,509 views, 2026-10-30 https://x.com/Asteri_eth/status/2116111471834743106; @twoclipping (9,487 followers), 3,801 saves / 2,194 likes / 143,709 views, 2026-09-26 https://x.com/twoclipping/status/2103835273813496100; varg how to skill "Beat-Synced Cut" https://varg.ai/how-to/skills/beat-synced-cut ### 04.3 Do sound design as its own pass: map events, prototype three dry SFX, lay room tone under every edit, mix dialogue first to numeric levels. Wrong sounds come from leaving audio to the video model, so foley and sound editing are separate. An agent can't hear, so it maps events, checks peaks and mono, mixes dialogue first to measured levels with room tone and no dead silence, and asks the human which prototypes work. Do: - Treat foley and sound editing as a pass separate from video generation. - Map the visual events by function (micro-cut, object lock, texture movement, transition, resolve) and start with three distinct dry sounds. - Export 48 kHz / 24-bit WAVs, a level-matched audition reel and a cue sheet; check peaks, clean endings and mono compatibility. - Mix dialogue first: speech around -18 to -12 dBFS (peaks -6 to -3), music under dialogue -30 to -20 dBFS, important SFX -18 to -8 dBFS. - Lay room tone/ambience at about -35 to -24 dBFS under every edit, never leave dead silence, and keep the master at or below -1 dBTP true peak. - Render a frames + spectrogram + sync-marker sheet to check timing, then ask the human which prototype works and iterate. Done when: Dialogue sits at -18 to -12 dBFS (or -16 to -14 LUFS) with music well under it, no edit has dead silence, the master never exceeds -1 dBTP, and the human has picked the working SFX prototypes. ```text [Build an original SFX palette for my short visual edit in Codex, without a DAW or third-party samples. First map the visual events by function: micro-cut, object lock, texture movement, transition, resolve. Start with only three distinct dry mechanical-digital sounds, not a full pack. Avoid tonal beeps, pitch-sweep lasers, bells, conventional drums and long reverb. Export individual 48 kHz / 24-bit WAVs, a level-matched audition reel and a cue sheet. Check peaks, clean endings and mono compatibility. Don’t claim you can hear the result: ask me which prototype works, which fails and why; iterate from that feedback before making variations.] ``` ```text Act as a professional film sound mixer and fully mix this Premiere Pro sequence. Use these rules: Dialogue: always priority. Normal speech around -18 to -12 dBFS, with peaks generally -6 to -3 dBFS. Dialogue consistency: clip-gain first, then light compression. Avoid obvious pumping. Music under dialogue: typically -30 to -20 dBFS depending on density. Duck it enough that every word is effortless to understand. Music without dialogue: can rise to roughly -18 to -10 dBFS. Ambience/room tone: roughly -35 to -24 dBFS. Never leave unnatural dead silence between edits. Background Foley: usually -30 to -18 dBFS. Important Foley/SFX: roughly -18 to -8 dBFS. Major cinematic impacts, gunshots, explosions, transitions: may peak around -6 to -2 dBFS, but never clip. Master: never exceed -1 dBTP true peak. ``` Proof: @AmirMushich (91,808 followers), 959 saves / 668 likes / 43,099 views, 2026-09-27 https://x.com/AmirMushich/status/2104257004474671129; @beginnersblog1 (29,874 followers), 93 saves / 81 likes / 3,843 views, 2026-10-29 https://x.com/beginnersblog1/status/2115772390890438852; varg how to skill "OpenMontage Sound Design Rules" https://varg.ai/how-to/skills/openmontage-sound-design; @Asteri_eth (3,340 followers), 21 saves / 26 likes / 631 views, 2026-11-03 https://x.com/Asteri_eth/status/2117629600901534024; varg how to skill "How to make video sound effects in code with earshot" https://varg.ai/how-to/skills/earshot; @Damir_Akaza (5,278 followers), 21 saves / 40 likes / 1,475 views, 2026-11-01 https://x.com/Damir_Akaza/status/2117000237025804546; @beechinour (18,923 followers), 1,952 saves / 1,042 likes / 104,679 views, 2026-09-07 https://x.com/beechinour/status/2097036762283458575 Top 10: 1. Use bodily actions (e.g. sniffles between lines) as punctuation in dialogue; the author found this the biggest improvement. (https://x.com/second_bassoon/status/2117307653320007954) 2. For speaking clips, generate the voice separately (ElevenLabs or a recording), then paste the exact script of that audio into the video prompt. (https://x.com/ZssBecker/status/2113320998527221875) 3. Mix dialogue first: normal speech around -18 to -12 dBFS with peaks around -6 to -3 dBFS. (https://x.com/beechinour/status/2097036762283458575) 4. Lay room tone and ambience under every edit at roughly -35 to -24 dBFS, and never leave unnatural dead silence between edits. (https://x.com/beechinour/status/2097036762283458575) 5. Put scene cuts and title entrances on the beat; if the animation can't change, move where the music starts. (https://x.com/Voxyz_ai/status/2109358372873183564) 6. Before making SFX, map the visual events by function: micro-cut, object lock, texture movement, transition, resolve. (https://x.com/AmirMushich/status/2104257004474671129) 7. Cut reference audio to exact whole seconds (6, 8, 10, 15) with no leftover tails before uploading it. (https://x.com/Kiber_Alla/status/2078814921756754352) 8. When a song exists, name the uploaded audio as the master timing source and tell the model to sync cuts, breath, gaze and gestures to it. (https://x.com/Israa_Ali2077/status/2117286613739032729) 9. Record the narration first, get word timestamps from it, and let the voice timing drive every cut. (https://varg.ai/how-to/prompts/x-2102844654169575547) 10. Write a rhythm sheet before animating: a hook inside 3 s, a burst in the middle, and one beat of silence before the biggest hit. (https://x.com/Asteri_eth/status/2117328625552826530) ## 05. Characters, props, details A turnaround sheet for every character, a sheet for every hero object. ### 05.1 Build one reference sheet per character look, then give the video model only cropped clean panels with one large face. Sources show a sheet with profile, matching-scale turnaround, face close-ups and expressions as the source of truth, but small faces inside dense full-body views make video models drift; one sheet per look and 2–3 cropped panels avoid it. Do: - From the reference photo, generate a continuity sheet: character profile, full-body turnaround (front, 3/4, side, back) at matching scale, face close-ups, expression row. - Make a separate sheet for every look: a face close-up plus full-body front and back on plain grey. - Remove or block the face in full-body views so the macro close-up is the only face the model learns from. - Crop the hero portrait and the front full-body (sometimes 3/4) into separate clean reference files for video generation. - Place the character into each location as a still keyframe and fix it before animating. Done when: Every look has its own sheet, each shot gets 2–3 cropped panels with a single large face, and the keyframe in each location looks right before any clip is generated. ```text Create a professional AI Video Production — Character Reference Sheet for the provided character image. Create a polished 4:5 portrait continuity sheet focused exclusively on the character: 1. CHARACTER PROFILE — name, role, approximate age, height, body type, personality, distinctive traits, and signature colors. 2. FULL-BODY TURNAROUND — front, 3/4, side, and back views of the exact same character at matching scale, with identical facial features, hairstyle, body proportions, costume, colors, and accessories. 3. FACE AND IDENTITY DETAILS — front, profile, and 3/4 head-and-shoulder close-ups showing facial structure, eyes, eyebrows, nose, lips, hairstyle, skin tone, makeup, scars, tattoos, and other defining features. 4. EXPRESSION SHEET — neutral, happy, angry, sad, surprised, worried, confident, and determined expressions. ``` ```text Character turnaround sheet of the person from the reference photo, six views: full body front, side and back + close-up front, side and back. Strictly keep 100% of face, hair, body, clothing and all details from the reference. Clean light gray background, consistent and natural lighting. ``` ```text Film character sheet, photorealistic commercial studio photography. Three views of the exact same person from the attached reference image, standing in a row on a seamless neutral grey #B8B8B8 background: full-body front view, full-body back view, and a macro close-up portrait of the face. Identical person, outfit, hairstyle and styling in all three views, matched true to the reference photo. ``` ```text Character reference sheet for [name], [age, face, hair, outfit]. Three panels: close-up, full body front, full body back. ``` Proof: @OlatundeAI (6,183 followers), 6,292 saves / 3,323 likes / 336,821 views, 2026-08-20 https://x.com/OlatundeAI/status/2090374807753728011; @meAsifAi (4,494 followers), 3,450 saves / 1,834 likes / 104,416 views, 2026-09-10 https://x.com/meAsifAi/status/2098429368053960936; @meAsifAi (4,494 followers), 1,191 saves / 724 likes / 35,711 views, 2026-08-25 https://x.com/meAsifAi/status/2092365498134696162; @diamai_ (1,065 followers), 31 saves / 49 likes / 1,008 views, 2026-10-28 https://x.com/diamai_/status/2115372779570704540; @simonecanciello (28,414 followers), 4,027 saves / 2,317 likes / 338,967 views, 2026-10-30 https://x.com/simonecanciello/status/2116268465463976312; @meAsifAi (4,494 followers), 3,450 saves / 1,834 likes / 104,416 views, 2026-09-10 https://x.com/meAsifAi/status/2098364205980356955 ### 05.2 Build a sheet for every hero object from real photos, name every part to keep, and give it its own reference slot and lock line. Sources say a product sheet keeps the product consistent across angles and text, real photos keep packaging accurate, naming each detail (check pattern, logo placement, a car's silhouette and weapon mounts) helps it survive, and props on moving characters melt unless kept small and still. Do: - Upload the real product photo, generate a product sheet, and pair it with a per-second storyboard before generating video. - In the prompt list every part to preserve: pattern, collar, pocket alignment, logo placement and texture, or a vehicle's silhouette, blades, wheels, lights and weapon mounts. - Give each hero object its own reference image slot, mapped by the actual upload order. - On the character sheet, show the character holding the main prop and add prop and equipment callouts. - In every shot prompt list recurring props in the consistency line and add a STRICT LOCK per hero prop: never disappears, falls off, changes shape or gets replaced. - Keep swapped props small and still in shots where the character moves. Done when: Each hero object has a sheet or reference image in its own slot, its parts to preserve are named in the prompt, and every shot prompt carries its lock line. ```text Create a high-end cinematic product sheet + storyboard board in 16:9 landscape format for a 15-second premium fashion spec commercial featuring reference shirt, a refined red-and-white micro-check long-sleeve button-up shirt. Use Image A / reference shirt as the exact product reference and keep the shirt highly accurate and consistent throughout the full board, including its small red-and-white check pattern, standard collar, front button placket, long sleeves, curved hem, and left chest pocket with diagonal check alignment. ``` ```text REFERENCE ASSIGNMENTS — FOLLOW THE ACTUAL ATTACHMENT ORDER Reference 1: HERO CAR — red-and-black HT-768 KURENAI. Preserve its low armored silhouette, horn-like front blades, wheels, red illumination and existing weapon mounts. Reference 2: HERO PILOT — the red, gold and black mechanical samurai with the horned helmet. This character drives ``` ```text STRICT HEADPHONE LOCK: From the very first frame to the final frame, she wears large, matte-black over-ear headphones covering both ears. Both ear cups and the headband remain clearly visible and securely positioned over her head throughout every punch, jump, kick, fall, and rapid camera movement. Her headphones never disappear, fall off, change shape, or get replaced. Her hair moves naturally around the headphones. Do not use tiny earbuds. ``` Proof: varg how to prompt "Product sheet + 15-frame storyboard, then a shirt commercial" https://varg.ai/how-to/prompts/x-2098058251921457332; @0xForce_ (7,601 followers), 39 saves / 47 likes / 1,633 views, 2026-11-03 https://x.com/0xForce_/status/2117412908413944290; @theodorsan (649 followers), 375 saves / 303 likes / 18,727 views, 2026-10-30 https://x.com/theodorsan/status/2116199275105275984; @Just_sharon7 (48,986 followers), 491 saves / 377 likes / 279,066 views, 2026-07-23 https://x.com/Just_sharon7/status/2080193844776251575; @tokyo_Valentine (11,032 followers), 28 saves / 21 likes / 1,128 views, 2026-10-29 https://x.com/tokyo_Valentine/status/2115783746083602528; @oggii_0 (18,349 followers), 83 saves / 167 likes / 19,342 views, 2026-09-03 https://x.com/oggii_0/status/2095410797837226221; @meAsifAi (4,494 followers), 306 saves / 285 likes / 11,917 views, 2026-09-15 https://x.com/meAsifAi/status/2099828931889144192; @MissDelulu9 (14,729 followers), 135 saves / 205 likes / 8,855 views, 2026-07-12 https://x.com/MissDelulu9/status/2076276747038060849; @HeyAbhishek (111,373 followers), 347 saves / 330 likes / 20,004 views, 2026-10-28 https://x.com/HeyAbhishek/status/2115429480793874676; @MissDelulu9 (14,729 followers), 20 saves / 158 likes / 7,526 views, 2026-10-28 https://x.com/MissDelulu9/status/2115433159160672383; @denogrowth (2,841 followers), 263 saves / 245 likes / 605,772 views, 2026-11-02 https://x.com/denogrowth/status/2117401595612700860; varg how to skill "Turn product photos into AI video that keeps the logo and shape" https://varg.ai/how-to/skills/consistent-product-ai-video ### 05.3 Give every reference a scoped role in the prompt and add a lock block that forbids redesign, layout copying and extra figures. Sources show models copying sheet backdrops, poses and studio light, reproducing side-by-side layouts and adding extra people; prompts that say what each image defines, label refs look-only or set-only, list immutable facts, say do not redesign and repeat verbatim in every clip keep identity. Do: - Write a REFERENCE ROLES list: which image defines each character's identity and wardrobe, which defines the location, which is the hero object, mapped by upload order. - Label each reference look only or set only, and say what not to take from it: grey studio backdrop, standing pose, flat studio lighting, original background. - Add a CHARACTER CONSISTENCY block naming the reference as the exact visual identity, ending with Do not redesign her. - List immutable facts with numbers (height, build) and side-specific details such as LEFT wrist or LEFT ear. - State the exact number of characters and forbid duplicates and the reference's side-by-side layout. - Paste the same locked blocks verbatim into every single clip prompt. Done when: Each shot prompt names every reference's role and exclusions, its lock block covers face, hair, outfit, props, head count and layout, and the block text is identical in every clip. ```text REFERENCE ROLES @ Image1: identity, face, hair and wardrobe of the HEROINE. @ Image2: identity and wardrobe of ENEMY 1, ENEMY 2 and ENEMY 3. @ Image3: location layout, buildings, debris and lighting. ``` ```text Refs (make first with gpt_image_2_5): @hero_sheet = {HERO: e.g. shepherd-dog knight, tan and black fur, worn leather armour} look only. @partner_sheet = {PARTNER: e.g. black panther rogue, purple scarf} look only. @town, @forest = location stills, set only. ``` ```text @Image 1 defines the woman's face, her split black-and-silver shaggy shoulder- length hair, and her wardrobe only — black ribbed tank top with a buckled eyelet strap over the right shoulder, spiked cross-stud choker, silver cross pendant, black chained arm warmers, studded belt, baggy black cargo pants, chunky black boots. Do not copy the grey studio backdrop, the standing pose, or the flat studio lighting from @Image 1. ``` ```text CHARACTER CONSISTENCY: Use the provided Natalia character reference as the exact visual identity. Preserve her facial features, face shape, natural hair color, skin tone, and overall appearance. Keep her approved hairstyle, makeup, dark Joseon-inspired assassin outfit, sword, and accessories exactly consistent throughout every shot. Do not redesign her. ``` Proof: @AiwithBloodline (7,256 followers), 80 saves / 128 likes / 5,892 views, 2026-10-29 https://x.com/AiwithBloodline/status/2115841225106604469; @iamsofiaijaz (15,879 followers), 76 saves / 136 likes / 6,157 views, 2026-10-28 https://x.com/iamsofiaijaz/status/2115406342802506033; @oggii_0 (18,349 followers), 83 saves / 167 likes / 19,342 views, 2026-09-03 https://x.com/oggii_0/status/2095410797837226221; varg how to prompt "Animal-knight fantasy short with dialogue, 21:9" https://varg.ai/how-to/prompts/x-2116728095323677043; @AIwithNatalia (12,525 followers), 25 saves / 212 likes / 24,404 views, 2026-11-01 https://x.com/AIwithNatalia/status/2116887713563787586; @abulu8 (5,176 followers), 39 saves / 147 likes / 14,928 views, 2026-11-01 https://x.com/abulu8/status/2116999082413904166; @abulu8 (5,176 followers), 49 saves / 155 likes / 58,443 views, 2026-10-31 https://x.com/abulu8/status/2116476626917847495; @ChillaiKalan__ (16,607 followers), 40 saves / 101 likes / 12,937 views, 2026-11-01 https://x.com/ChillaiKalan__/status/2116775135067689114; @Mavern_0x (3,413 followers), 14 saves / 48 likes / 2,662 views, 2026-11-02 https://x.com/Mavern_0x/status/2117403866194903400; @meAsifAi (4,494 followers), 3,450 saves / 1,834 likes / 104,416 views, 2026-09-10 https://x.com/meAsifAi/status/2097949170594193427; varg how to prompt "Whiteboard explainer ad, 9 clips chained by last frame" https://varg.ai/how-to/prompts/x-2086118407708250149 Top 10: 1. Draw a full-body turnaround (front, 3/4, side, back) of the exact same character at matching scale, with identical face, hair, proportions, costume, colours and accessories. (https://x.com/OlatundeAI/status/2090374807753728011) 2. Keep the full board as the design lock, but feed the video model only 2–3 cropped panels (hero portrait, front full-body, sometimes 3/4) as separate clean references. (https://x.com/meAsifAi/status/2098429368053960936) 3. Lock the character first with one clean character sheet and reference it in every generation; this is what stops the drift. (https://x.com/_VVSVS/status/2090813096310784160) 4. Do not leave a small face inside the full-body views; video models lock onto that tiny blurry face and the character drifts mid-generation. (https://x.com/meAsifAi/status/2092365498134696162) 5. Keep swapped props small and still; detailed items like shoes melt into blur the moment the character walks. (https://x.com/denogrowth/status/2117401595612700860) 6. When a stylized reference becomes live action, ask for real skin, hair, fabric and materials: natural pores, individual hair strands, clothing seams and folds. (https://x.com/ZentrixHQ/status/2116844865279451613) 7. Pair a product sheet with a per-second storyboard so fine patterns survive macro camera moves. (https://varg.ai/how-to/prompts/x-2098058251921457332) 8. Scope each reference to what it defines (face, hair, wardrobe only) and forbid copying the sheet's grey backdrop, standing pose and flat studio lighting. (https://x.com/AiwithBloodline/status/2115841225106604469) 9. For several outfits, label each outfit panel on the sheet, tie each to its scene, and forbid redesigning, mixing or substituting outfits. (https://x.com/sophiaparkerr_/status/2116891573799764118) 10. Generate a six-view turnaround from a reference photo (full body front, side, back plus close-up front, side, back), keeping 100% of face, hair, body and clothing. (https://x.com/simonecanciello/status/2116268465463976312) ## 06. Blockout in code Build the world in three.js or Blender, with your reference painting and generated images as textures. ### 06.1 Render a still of the coded build beside the reference picture and fix the differences until they line up, before any animation. Authors building three.js machines from one picture had the agent render a still beside it and close the gaps for 3 or 4 rounds before filming. Another scores the render against the reference with a script instead of eyeballing it. Others compare front and side views with the reference. Do: - Let the reference picture set the look and build every part as three.js code, or as Blender Python through MCP for Blender. - Block out the silhouette and proportions first, then build the real geometry, then add details and materials. - Render one still beside the reference picture and patch the differences round after round before any animation. - Have the agent write a script that scores the render against the reference (color, light, contrast, detail) from 0 to 100; loop render, score, side-by-side, change one thing. - Render front, side and rear views, inspect the mesh for seams and overlaps, and compare front and side views with the reference. - Once it moves, review a grid of frames at every beat, fix what looks off and re-render until a full pass comes back clean; save the .blend and the .glb. Done when: The agent's own still and the reference picture line up, and a full frame-grid pass comes back clean. ```text > Read the scene and open a fresh collection. > Block out the silhouette and proportions first. > Build the real geometry with Blender tools and bpy. > Add the details and the materials. > Render front, side and rear. > Compare, fix what's wrong, save the .blend and the .glb. ``` Proof: @0x_fokki (16,348 followers), 7,340 saves / 6,751 likes / 661,794 views, 2026-10-30 https://x.com/0x_fokki/status/2116289885870981296; @elvar_official (40,323 followers), 4,469 saves / 3,964 likes / 340,386 views, 2026-10-30 https://x.com/elvar_official/status/2116192984299344372; @0x_fokki (16,348 followers), 225 saves / 212 likes / 22,910 views, 2026-11-01 https://x.com/0x_fokki/status/2116968438296150116; @rewind02 (7,898 followers), 664 saves / 384 likes / 45,229 views, 2026-11-01 https://x.com/rewind02/status/2116918597658305001; @rewind02 (7,898 followers), 438 saves / 323 likes / 41,707 views, 2026-10-30 https://x.com/rewind02/status/2116209421848064265; @sorenxet (1,035 followers), 80 saves / 66 likes / 2,195 views, 2026-11-03 https://x.com/sorenxet/status/2117608142200512621; @beamnxw (14,694 followers), 1,194 saves / 881 likes / 69,200 views, 2026-10-22 https://x.com/beamnxw/status/2113275303808160226; varg how to skill "Teardown by Elvar", 302 stars https://varg.ai/how-to/skills/elvar-teardown; varg how to skill "img2threejs by Nick (Hoài Nhớ)", 19,931 stars https://varg.ai/how-to/skills/img2threejs; @rewind02 (7,898 followers), 424 saves / 350 likes / 68,995 views, 2026-10-23 https://x.com/rewind02/status/2113573386794254715; @danikkk_wqs (1,502 followers), 14 saves / 17 likes / 1,021 views, 2026-11-03 https://x.com/danikkk_wqs/status/2117570805999521960 ### 06.2 Solve staging, facing and camera in a grey-box previz, feed its clip and a start frame to the video model, then check it followed. A 3D previz rendered by Seedance kept the motion and camera control. Seedance sometimes showed too few characters until one author built a previz with Claude. A Blender greybox clip as video reference landed 6 hard cuts within 0.07 s of plan. Seedance still isn't perfect at matching, so check. Do: - Give the agent the scene description and references and build a simple grey-box previz in three.js or Blender. - Specify the shots, camera moves and final pose, and work out where everyone stands, which way they face and where the camera goes. - Generate a start frame, block out the scene in Blender and animate the camera, then render the greybox clip per shot. - Feed the video model the start frame, the camera clip and a prompt per shot, telling it to keep the frame and follow the camera movement exactly. - Where motion matters, pass a depth map as the motion reference and the character sheet as a separate identity reference. - Check whether the video follows the previz; keep what works, regenerate only what doesn't, and stop spending on an approach that keeps failing. Done when: Each generated shot keeps the previz layout (who stands where, facing, camera), and cuts land on the planned times. ```text Use [img1] as the clean start frame. Don't change anything about the frame. Use [vid1] as a camera motion reference. Follow the camera movement in the reference exactly, including the slow push in on the man, and the fast pan left to the other man. You can move the characters. ``` ```text THE CAUSEWAY: PULL BACK FROM THE MAKI TO THE DRAGONS 6WHM Audio is strictly limited to natural diegetic sound effects and environmental sound, with absolutely no music, score, soundtrack, background music, musical cues, melodic tones, rhythmic beats, or instrumental audio of any kind. SCENE CONTEXT: A 15-second photorealistic live-action shot from a big-budget feature film, one continuous handheld take filmed for real on location by a camera operator moving backwards down a course. It opens on a tight close-up of a small, terrified Maki on a start column, looking up into the lens. ``` Proof: @venturetwins (249,386 followers), 3,486 saves / 6,074 likes / 732,503 views, 2026-06-24 https://x.com/venturetwins/status/2069809200788799582; @reidhannaford (3,398 followers), 1,160 saves / 1,625 likes / 229,667 views, 2026-06-29 https://x.com/reidhannaford/status/2071595581508563168; @reidhannaford (3,398 followers), 260 saves / 352 likes / 17,790 views, 2026-06-30 https://x.com/reidhannaford/status/2071956990847295855; @reidhannaford (3,398 followers), 260 saves / 352 likes / 17,790 views, 2026-06-30 https://x.com/reidhannaford/status/2072041016979284252; @NEXUS_TO_NOVA (31,066 followers), 7 saves / 25 likes / 1,008 views, 2026-11-03 https://x.com/NEXUS_TO_NOVA/status/2117632582514864494; varg how to skill "Control AI Video Cuts and Camera With a Blender Greybox" https://varg.ai/how-to/skills/blender-greybox-ai-video-cuts; varg how to skill "Open-Source Previs Studio for AI Video Camera Control" https://varg.ai/how-to/skills/cozyclay-previs; varg how to component "varg Previz" https://varg.ai/how-to/component/varg.previz; varg how to skill "Previz to Seedance" https://varg.ai/how-to/skills/varg-previz; @beamnxw (14,694 followers), 1,194 saves / 881 likes / 69,200 views, 2026-10-22 https://x.com/beamnxw/status/2113275303808160226; @NexlowX (1,536 followers), 868 saves / 691 likes / 120,534 views, 2026-08-31 https://x.com/NexlowX/status/2094494202587398616; @PJaccetturo (101,079 followers), 28 saves / 27 likes / 8,001 views, 2026-10-30 https://x.com/PJaccetturo/status/2116160709180723486; varg how to prompt "Blender previz as the camera path: a 15 s one-take pull-back" https://varg.ai/how-to/prompts/x-2105285283298115670 ### 06.3 Texture coded structures from generated images (triplanar grids or camera projection), or layer a painting by depth, and keep reusing the same saved style references. StableGen projects images generated around a model onto it, each view referencing earlier ones, so the texture stays consistent. A Godot build textured from Nano Banana 2x2 grids with triplanar mapping, no UV unwrapping. One street shot came from a single 4K painting split into depth layers. Do: - Collect mood references, generate style reference images, and save every prompt that works as JSON. - Build buildings and big structures directly in code; write prompts for groups of 10 related props and make one reference image per group for image-to-3D. - Generate textures as 2x2 grids, crop them into 4 images, derive normal and roughness from the colour, and apply them with triplanar mapping. - For a hero model, place cameras around it, write a texture prompt and project the generated views onto the surface. - For a painted set, split one 4K painting into layers with a depth map and render the camera move in code. - Generate model textures with baked lighting removed so the scene can relight them, then prompt follow-ups for textures, lighting and alpha masking. Done when: The texture stays consistent all the way round each model, and the style does not drift between assets. Proof: @hasantoxr (436,534 followers), 987 saves / 650 likes / 35,416 views, 2026-10-31 https://x.com/hasantoxr/status/2116457666403029040; @PatiChanchi (472 followers), 1 saves / 21 likes / 160 views, 2026-10-31 https://x.com/PatiChanchi/status/2116660513094660144; @rewind02 (7,898 followers), 42 saves / 71 likes / 2,270 views, 2026-11-03 https://x.com/rewind02/status/2117586491350135052; @doodi_dev (230 followers), 134 saves / 104 likes / 4,809 views, 2026-11-03 https://x.com/doodi_dev/status/2117440376089010336; @sorenxet (1,035 followers), 69 saves / 92 likes / 5,353 views, 2026-11-02 https://x.com/sorenxet/status/2117279394540519635; @rewind02 (7,898 followers), 64 saves / 90 likes / 3,436 views, 2026-11-02 https://x.com/rewind02/status/2117278100480262484; @MengTo (190,284 followers), 2,151 saves / 1,688 likes / 618,010 views, 2026-08-07 https://x.com/MengTo/status/2085765403729653877 Top 10: 1. Render one still next to the reference picture and fix the gaps for about 3 rounds before animating anything. (https://x.com/0x_fokki/status/2116289885870981296) 2. Build a 3D previz, then have the video model render the final shots from it so the motion and camera control are preserved. (https://x.com/venturetwins/status/2069809200788799582) 3. Solve the scene in a previz first: test staging, timing and camera movement before you ever touch the video model. (https://x.com/reidhannaford/status/2071595581508563168) 4. Generate a start frame, block the scene out in Blender with an animated camera, then feed both the frame and the camera clip to Seedance. (https://x.com/reidhannaford/status/2071956990847295855) 5. Give a director's brief first: specify the shots, expressions, camera moves and final pose before the scene is built. (https://x.com/beamnxw/status/2113275303808160226) 6. Texture a model by generating images around it and projecting them onto it; for existing models, place cameras, write a texture prompt and generate. (https://x.com/hasantoxr/status/2116457666403029040) 7. Feed the video model a depth map instead of raw footage as the motion reference, so motion is the only thing left to interpret. (https://x.com/NexlowX/status/2094494202587398616) 8. Use a Blender previz video as the camera-path reference to steer a complex one-take camera move in the video model. (https://varg.ai/how-to/prompts/x-2105285283298115670) 9. Establish silhouette and proportions first, then build real geometry with Blender tools and Python, then details and materials. (https://x.com/rewind02/status/2116918597658305001) 10. Run a closed loop: the agent writes code, runs it, looks at the rendered result and fixes what it sees. (https://x.com/QibazX/status/2115490139631743316) ## 07. Takes Animate each shot in code; use a video model only for the shots code can't do. Make several takes, keep the best. ### 07.1 Tag every shot code or model: render graphic, UI and stylized shots in code; send only real-motion shots to a video model, from a keyframe. Code shots (three.js, Remotion, HyperFrames, Manim) stay consistent and editable and render straight to MP4. Video models are kept for motion shots made from image-model keyframes; Seedance does better with motion shots. Marketing skills start by picking code, a video model or an avatar. Do: - Decide per video and per shot whether it is code, a video model or an avatar before scripting and prompting. - Render motion-graphics, map and explainer shots with HyperFrames, Remotion, Manim or Motion Canvas instead of a video model. - Render stylized cartoon shots (2D characters in a 3D world) in three.js and Remotion so every shot stays consistent and editable. - For each remaining model shot, make a keyframe with an image model first, then generate with Seedance 2.5 or Veo 3.1; prefer Seedance for motion shots. - In hybrid pieces, let code-built diagrams act out the content while a video model performs only the recurring host. - If one shot needs readable phone or screen text, switch models for that shot or blur the screen and composite the real screen later. Done when: Every shot in the plan is marked code or model, and every model shot has a keyframe still before any generation runs. ```text Using the varg-ai and hyperframes skills, make a 2-minute 16:9 edu-music video: a recurring host sings a lesson on {TOPIC} while code-built diagrams act out each line. ``` Proof: @RoundtableSpace (271,067 followers), 849 saves / 533 likes / 68,917 views, 2026-11-02 https://x.com/RoundtableSpace/status/2117236532482396349; @deedydas (261,658 followers), 4,381 saves / 2,099 likes / 159,803 views, 2026-09-29 https://x.com/deedydas/status/2104957026199900220; varg how to style "Code-made motion design" https://varg.ai/how-to/styles/trending; @rexan_wong (24,918 followers), 18,188 saves / 6,888 likes / 647,682 views, 2026-09-26 https://x.com/rexan_wong/status/2103707054108299437; varg how to skill "Marketing Video Production", 86,225 installs, 56,690 stars https://varg.ai/how-to/skills/marketingskills-video; varg how to prompt "Agent-made explainer song: host sings rhythm math" https://varg.ai/how-to/prompts/x-2112601639672566182; varg how to prompt "Heist-movie reply ad to a copycat restaurant" https://varg.ai/how-to/prompts/x-2083364450900549639; @maxt3chno (1,375 followers), 1,789 saves / 1,193 likes / 62,543 views, 2026-10-28 https://x.com/maxt3chno/status/2115386626775953881 ### 07.2 Iterate on stills first, then animate the approved start frame with role-tagged inputs and a motion-only, timestamped prompt; lock audio for speech. Images are cheap and video is expensive, so iterate on stills. The sheet, start frame, video chain gives the saved found-footage look. Role tags stop the model guessing; re-describing the scene in image-to-video melts the background; locked audio stops Seedance adding moves. Do: - Iterate on the character sheet, location images and the exact start frame before any video, because images are cheap and video is expensive. - Build each model shot as character sheet, then start-frame still, then video; say the start frame matches the reference exactly. - Bind every input to a role with tags like , , , , and say what each reference must not lend (backdrop, pose, people, text). - When animating the approved still, prompt only the motion and camera move, timed in brackets like [0-3s]; do not re-describe the scene. - For speaking shots, generate or record the voice first, attach it as the audio reference and paste the exact script into the prompt. - For sequences longer than one generation, chain by last frame with locked style blocks pasted verbatim, and check the joins, since stitching can break camera movement and pace. Done when: Each model shot has an approved start-frame still, every input bound to a role, a motion-only timed prompt, and a locked audio file if it speaks. ```text persuasive UGC creator, expressive gestures, says exactly: {your script}. casual unpolished single-take phone video, full eye contact, start frame matches the reference exactly, no objects, no cuts, no music. ``` ```text Refer to Video 1 only for camera movement and pacing; do not copy its people, places or text. ``` ```text [0-3s] Person starts walking [3-6s] They stop and look at the camera ``` Proof: @EXM7777 (156,736 followers), 805 saves / 440 likes / 29,166 views, 2026-09-17 https://x.com/EXM7777/status/2100615302388330944; varg how to prompt "Y2K home-video road trip: who holds the camera is the prompt" https://varg.ai/how-to/prompts/x-2102775755646337530; varg how to prompt "Persuasive talking head from a Pinterest frame" https://varg.ai/how-to/prompts/x-2099547856486989951; @_philschmid (126,324 followers), 484 saves / 355 likes / 21,810 views, 2026-08-27 https://x.com/_philschmid/status/2093012878211072183; @AiwithBloodline (7,256 followers), 80 saves / 128 likes / 5,892 views, 2026-10-29 https://x.com/AiwithBloodline/status/2115841225106604469; varg how to prompt "Heist-movie reply ad to a copycat restaurant" https://varg.ai/how-to/prompts/x-2083364450900549639; @imrollandex (6,349 followers), 283 saves / 228 likes / 21,684 views, 2026-08-28 https://x.com/imrollandex/status/2093367790241464609; @ZssBecker (1,325,986 followers), 3,135 saves / 1,050 likes / 136,548 views, 2026-10-22 https://x.com/ZssBecker/status/2113320998527221875; varg how to prompt "Whiteboard explainer ad, 9 clips chained by last frame" https://varg.ai/how-to/prompts/x-2086118407708250149; @laszlogaal_ (2,746 followers), 241 saves / 319 likes / 20,938 views, 2026-06-18 https://x.com/laszlogaal_/status/2067543462178931145 ### 07.3 Make several cheap takes per shot, review frames by severity, keep only clean seconds, and finish only the keeper. One 22-minute film kept 127 of 847 takes; a stunt test needed 4 Seedance and 7 Omni Flash attempts. Rerolling at 480p and upscaling only the keeper saves credits. Stills beside the reference, frame-grid reviews and code checks catch faults; trial reels show which variant viewers keep watching. Do: - Draft model takes at 480p, evolving the prompt, and render or upscale only the keeper at final resolution. - Budget several attempts per hard shot and keep the best one; expect most takes to be discarded. - For code shots, render one still next to the reference and fix the gaps for several rounds before animating. - Review each render as a grid of frames per beat, list problems by severity without touching anything, run code checks (text size, blank or frozen frames, colours, foot sliding, loop seams), then fix and re-render until a pass is clean. - Inspect model takes for anatomy errors such as extra fingers and cut only the best seconds of each take, like live action. - When two variants both pass, post them as trial reels, check watch-time drop-off after 24 h and post the winner. Done when: A full review pass of every kept take comes back clean, and only the keeper of each shot has been finished at final resolution. Proof: @CharlieDFilm (4,171 followers), 6,063 saves / 6,040 likes / 1,271,505 views, 2026-09-13 https://x.com/CharlieDFilm/status/2098929162735997071; @YourAlphaMom (5,473 followers), 131 saves / 202 likes / 67,899 views, 2026-06-08 https://x.com/YourAlphaMom/status/2063952651205677170; @buraktuyan (2,274 followers), 530 saves / 2,015 likes / 680,185 views, 2026-08-24 https://x.com/buraktuyan/status/2092008235075084363; @thetripathi58 (114,295 followers), 135 saves / 205 likes / 113,960 views, 2026-10-17 https://x.com/thetripathi58/status/2111482598937755902; @0x_fokki (16,348 followers), 7,340 saves / 6,751 likes / 661,794 views, 2026-10-30 https://x.com/0x_fokki/status/2116289885870981296; @Voxyz_ai (27,685 followers), 1,440 saves / 711 likes / 130,399 views, 2026-09-27 https://x.com/Voxyz_ai/status/2104194355556671987; @shannholmberg (37,522 followers), 2,525 saves / 1,200 likes / 76,793 views, 2026-10-19 https://x.com/shannholmberg/status/2112104872451756348; @theparuchh (1,881 followers), 197 saves / 132 likes / 8,290 views, 2026-10-30 https://x.com/theparuchh/status/2116244999096398074; varg how to prompt "30s breakdance montage from three pose stills" https://varg.ai/how-to/prompts/yt-kling4-breakdance-montage; varg how to skill "Make a short film or music video with Kling 4" https://varg.ai/how-to/skills/kling-4-keyframe-film; @juliapintar (18,347 followers), 267 saves / 164 likes / 17,861 views, 2026-10-18 https://x.com/juliapintar/status/2111864278303363506; @rincyXIII (160 followers), 903 saves / 554 likes / 36,962 views, 2026-10-31 https://x.com/rincyXIII/status/2116528121138586015 Top 10: 1. Render stylized animated shots in code (three.js, Remotion) instead of a video generator so every shot stays consistent and editable. (https://x.com/RoundtableSpace/status/2117236532482396349) 2. Pick 1-2 reference videos and tell the agent to match them; naming a style works better than describing it. (https://x.com/rexan_wong/status/2103707054108299437) 3. Go character sheet, then a start-frame still, then video, instead of straight text-to-video. (https://varg.ai/how-to/prompts/x-2102775755646337530) 4. When animating an approved still, prompt only the motion; re-describing the whole scene confuses the model and melts the background. (https://x.com/imrollandex/status/2093367790241464609) 5. Render one still next to the reference picture and fix the gaps over several rounds before animating anything. (https://x.com/0x_fokki/status/2116289885870981296) 6. Budget for many takes per shot: one 22-minute film generated 847 takes and kept 127 clips. (https://x.com/CharlieDFilm/status/2098929162735997071) 7. Bind every input to a role with tags such as , , and instead of letting the model guess. (https://x.com/_philschmid/status/2093012878211072183) 8. Write Seedance prompts as global style, characters, a timestamped storyboard, camera movement, negatives and ending constraints, saying what happens every few seconds. (https://x.com/maxxmalist/status/2083951871069487592) 9. Iterate the prompt with many cheap 480p test generations, then render the final versions at 720p. (https://x.com/buraktuyan/status/2092008235075084363) 10. For speaking shots, generate or record the voice first, attach it as the audio reference and put the exact script in the prompt. (https://x.com/ZssBecker/status/2113320998527221875) ## 08. Edit HyperFrames or Remotion: the timeline is code you can change with a sentence. ### 08.1 Keep the edit as small code files and turn each note into a small change instead of regenerating the video. Generated video starts from scratch on every prompt, so one small fix means a full re-roll. With one file per step, a note like 'title is a beat late' changes one line, you know which file to fix, and the rest stays the same. Do: - Write spec.md from a reference video you like: pacing, first motion, cut rhythm. - Pull real numbers into data.json with the source next to every figure, and write story.md (one point, one hero number, three beats) before anything moves. - Write every frame as code in Scene.tsx (Remotion) and keep the timing in timing.ts. - Apply each plain-language note as a small timing change and append every build, note and diff to an append-only log.md. - Save hand edits into the project files so the agent keeps them the next time it edits. - Before export, check off-beat cuts, stale numbers, missing sources and unused scenes; render stills, check them against the spec and data, and wait for approval. Done when: Each note changed only the file it targets, the four pre-export checks pass, and the stills were approved before final.mp4 was rendered. Proof: @francescoinweb3 (1,741 followers), 3,770 saves / 2,072 likes / 162,093 views, 2026-10-17 https://x.com/francescoinweb3/status/2111393917899808880; @francescoinweb3 (1,741 followers), 67 saves / 60 likes / 3,499 views, 2026-10-27 https://x.com/francescoinweb3/status/2115074284787798396; @Asteri_eth (3,340 followers), 52 saves / 68 likes / 3,329 views, 2026-10-29 https://x.com/Asteri_eth/status/2115758546067349910 ### 08.2 Edit footage from its transcript: get the cut list approved, cut silences, fillers and repeated takes, then check every cut before export. Editing off the actual transcript means the agent doesn't guess, and a ~12KB transcript plus PNG filmstrips replaces about 45M tokens of frames. The cut list waits for your yes, so you stay the director. Its grasp of multi-angle story is weak, so check those calls. Do: - Test on one short clip first, because transcription starts as soon as you run it. - Transcribe every word with timestamps, work from the transcript plus PNG filmstrips, and match transcripts to the story's beat sheet. - Propose the list of cuts in plain English and wait for OK before touching the cut or rendering. - Cut every silence over 0.6 seconds and every filler word, keep only the last take of a repeated sentence, and place the B-roll. - Add smooth audio fades between cuts to prevent pops, duck music under speech, and normalize audio to -14 LUFS. - Render a preview, check every cut with up to 3 fix passes, then export edit/final.mp4 and, if needed, a 60-second 9:16 cut. Done when: The cut list was approved before rendering, no silence over 0.6 s or repeated take is left, audio sits at -14 LUFS, and edit/final.mp4 came out within three fix passes. ```text Here's my raw video [file]. Transcribe it, then edit it with ffmpeg and Python, no editor: 1. Cut every silence over 0.6 seconds and every filler word 2. When I repeat a sentence, keep only the last take 3. Add a punch-in zoom on every key point 4. Burn in word-by-word captions, 2 lines max, low center 5. Open with a 3-second hook using the strongest line in the video 6. Normalize audio to -14 LUFS 7. Export a 16:9 cut for YouTube and a 60-second 9:16 cut for Shorts Show me the list of cuts before you render. ``` Proof: @Axel_bitblaze69 (128,822 followers), 2,316 saves / 1,050 likes / 117,505 views, 2026-07-18 https://x.com/Axel_bitblaze69/status/2078551452201754980; @plusxbt (12,928 followers), 40 saves / 58 likes / 1,049 views, 2026-11-03 https://x.com/plusxbt/status/2117629278640574646; @visionsbyjake (311 followers), 54 saves / 37 likes / 4,026 views, 2026-10-28 https://x.com/visionsbyjake/status/2115489752132853780; @RoundtableSpace (271,067 followers), 198 saves / 191 likes / 41,211 views, 2026-10-30 https://x.com/RoundtableSpace/status/2116149368629735425; @RoundtableSpace (271,067 followers), 85 saves / 92 likes / 48,743 views, 2026-11-01 https://x.com/RoundtableSpace/status/2116813746215481578; @RealMarvelX (2,102 followers), 16 saves / 43 likes / 1,973 views, 2026-10-31 https://x.com/RealMarvelX/status/2116452582575026302; @realMax0r (186,731 followers), 82 saves / 2,360 likes / 32,594 views, 2026-10-29 https://x.com/realMax0r/status/2115731377178149187; @realMax0r (186,731 followers), 82 saves / 2,360 likes / 32,594 views, 2026-10-29 https://x.com/realMax0r/status/2115859040819671192; varg how to skill "video-use", 29,979 stars https://varg.ai/how-to/skills/video-use ### 08.3 Measure the music first, then put every scene cut and title entrance on a verified beat, with one sound effect per cut. The agent can't hear, so it must measure every step instead of going by feel. librosa sometimes locks onto the off-beat; plotting the waveform below 120 Hz against the beat grid catches that before cuts are placed. Do: - Use a licensed track (yours, or one under Mixkit's free license) and note its source and license. - Find the beats with librosa, plot the waveform below 120 Hz against the beat grid, zoom into a few bars, and shift the whole grid if the kicks don't line up. - Plan the structure as bars at the track's tempo, for example 10 bars at 120 BPM, 2 seconds each. - Put scene cuts and title entrances on the beat; change the animation if you can, otherwise move where the music starts. - Give every cut and big move one ready-made sound effect file (whoosh, pop, click, typing), using the project's existing sound library first. - To build into a payoff, shorten shot lengths step by step: 3 s, 2 s, 1 s, half a second. Done when: On the plotted waveform the kick drums line up with the beat grid, and every scene cut and title entrance sits on a beat. ```text Add sound to this video. You can't hear, so measure every step and don't go by feel. Install ffmpeg and librosa first if they're missing. 1. Use licensed music: a track I give you, or one under Mixkit's free license. Note the source and the license. Don't synthesize the music or the sound effects in code. 2. Find the beats with librosa. It sometimes locks onto the off-beat, so plot the waveform below 120 Hz against the beat grid and zoom in on a few bars. If the kick drums don't line up, shift the whole grid. 3. Put the big moves, like scene cuts and titles appearing, on the beat. If you can change the animation, change it. If you can't, move where the music starts. 4. Give every cut and big move one ready-made sound effect file: whoosh for movement, pop for things appearing, click for clicks, typing sounds for typing. ``` Proof: @Voxyz_ai (27,685 followers), 1,498 saves / 825 likes / 111,314 views, 2026-10-11 https://x.com/Voxyz_ai/status/2109358372873183564; @twoclipping (9,487 followers), 1,428 saves / 617 likes / 54,141 views, 2026-09-23 https://x.com/twoclipping/status/2102554209166000267; varg how to skill "Rhythmic Cutting (3-2-1)" https://varg.ai/how-to/skills/rhythmic-cutting-3-2-1; @0xwhrrari (23,306 followers), 2,635 saves / 1,406 likes / 93,462 views, 2026-11-02 https://x.com/0xwhrrari/status/2117240315069276186 Top 10: 1. Give the agent 1-2 reference videos to match and name the style; without one it falls back to centered text, gradient backgrounds and fade-ins. (https://x.com/rexan_wong/status/2103707054108299437) 2. Drive edits with the HyperFrames CLI: scaffold, check, preview, snapshot, edit the timeline, render locally or in the cloud. (https://varg.ai/how-to/skills/hyperframes-cli) 3. Split the job into six small steps that each write one file (spec.md, data.json, story.md, Scene.tsx, timing.ts, final.mp4), each starting from the previous file. (https://x.com/francescoinweb3/status/2111393917899808880) 4. For footage edits, transcribe every word first and cut from the transcript instead of guessing from the video. (https://x.com/Axel_bitblaze69/status/2078551452201754980) 5. Save every correction (too choppy, music louder) as a lesson in the skill so each new edit lands closer to your style. (https://x.com/RoundtableSpace/status/2116149368629735425) 6. Put scene cuts and title entrances on the beat; change the animation if you can, otherwise move where the music starts. (https://x.com/Voxyz_ai/status/2109358372873183564) 7. Cut every silence over 0.6 seconds and every filler word, and when a sentence is repeated keep only the last take. (https://x.com/RoundtableSpace/status/2116813746215481578) 8. Build the video as editable code (scenes, timing, captions) instead of regenerating it, because with generation one small fix means a full re-roll. (https://x.com/francescoinweb3/status/2115074284787798396) 9. Measure whether the video parks: percent still frames, longest hold, motion, cuts per 10 s, plus 4K playback and loudness gates. (https://varg.ai/how-to/skills/scenario-motion-energy-check) 10. Write each shot as a pure function of time, frame(t), so the same t always gives the same picture. (https://x.com/Asteri_eth/status/2115758546067349910) ## 09. Self-review Measure the frames and the sound against the reference; don't let the model grade itself. ### 09.1 Review rendered frames against the reference with a separate critic, list problems before fixing, and re-render until a full pass comes back clean. The agent cannot watch footage, so it reviews frame grids at every beat. A builder approving its own work is the failure: the verifier sits in its own conversation, and a second model scores against a checklist built from your best work until it passes. Do: - Render a still beside the reference picture and patch the differences round after round before any animation. - Build a review scene: neutral floor, fixed camera, the reference pinned in the corner, clip name and number on screen; keep the same light and camera for every clip. - Hand the render to a review subagent in its own conversation that watches it second by second and lists problems by severity without touching anything. - Review a grid of frames at every beat, fix one thing at a time, and re-render. - Have a critic from a different model family score the work against a checklist built from your best work; below the bar, the output never reaches the human and gets redone. - Log every generation call with its cost and set a budget cap so the loop never runs away. Done when: A full review pass over the frame grids comes back clean and the critic's score clears the bar, so the human only sees work that already passed. ```text “read SHOTS.md. for each shot: build the camera, render a low-sample preview, check the first, middle and last frame, fix what’s off, then render final samples. log every change to NIGHT_LOG.md” ``` ```text step 1: the agent makes something... a post, an image, a video, a piece of software step 2: a second model reviews it and scores it against a checklist built from your best work step 3: if the score is too low -> the output never reaches you, the agent takes the notes and redoes it step 4: repeat until the score clears the bar ``` Proof: @thedelost (5,536 followers), 55 saves / 55 likes / 3,743 views, 2026-10-28 https://x.com/thedelost/status/2115455685664096730; @anshuc (17,494 followers), 1,246 saves / 686 likes / 57,634 views, 2026-07-27 https://x.com/anshuc/status/2081801979131818412; @EXM7777 (156,736 followers), 358 saves / 235 likes / 16,838 views, 2026-09-03 https://x.com/EXM7777/status/2095572281804394895; @Voxyz_ai (27,685 followers), 1,440 saves / 711 likes / 130,399 views, 2026-09-27 https://x.com/Voxyz_ai/status/2104194355556671987; @rewind02 (7,898 followers), 782 saves / 611 likes / 27,451 views, 2026-10-29 https://x.com/rewind02/status/2115779810932728259; @rewind02 (7,898 followers), 271 saves / 270 likes / 12,463 views, 2026-10-31 https://x.com/rewind02/status/2116554764641063296; @de1lymoon (3,948 followers), 501 saves / 249 likes / 25,228 views, 2026-10-29 https://x.com/de1lymoon/status/2115855375777800330; @Mnilax (10,079 followers), 737 saves / 642 likes / 90,605 views, 2026-10-31 https://x.com/Mnilax/status/2116583914684702837; @QibazX (1,076 followers), 30 saves / 22 likes / 2,167 views, 2026-10-29 https://x.com/QibazX/status/2115863777966977150; varg how to skill "Business Motion Film", 223 installs, 1,229 stars https://varg.ai/how-to/skills/business-motion-film; varg how to skill "SeeCut", 398 stars https://varg.ai/how-to/skills/seecut; varg how to skill "Watch (video analysis)", 19,340 stars https://varg.ai/how-to/skills/claude-video-watch; varg how to skill "Huashu Art Motion by Huashu", 5,855 stars https://varg.ai/how-to/skills/huashu-art-motion; varg how to skill "Video references by Refs" https://varg.ai/how-to/skills/refs-video-references ### 09.2 Run programmatic checks first, a calibrated judge second with specific questions, and humans last to calibrate the judge; the main model fixes flagged issues. Checking thousands of frames by eye eats time and tokens; moving repeated checks to code and a decision model cut review cost by 90%. One overall score hides where output breaks, and model-written reference answers left one eval measuring its own noise. Do: - In code, check text size, blank or frozen frames and colours against the brand rules, and measure whether the video parks: percent still frames, longest hold, cuts per 10 s. - Score the render against the reference film on cut timing, brightness, motion, loudness and format, and check each clip against its shot spec: dolly in, held angle, end frame. - Check the mix: dialogue −16 to −14 LUFS, music 18–20 dB under speech, whoosh timing and true-peak limits. - Ask the decision model specific yes/no questions about selected frames and score one number per dimension, not one total. - Calibrate the judge on a golden set of 50-100 examples that includes failures, to high-80s or 90s agreement with a human (Cohen's kappa). - Before changing a prompt or model, replay past runs and diff the results; send flagged frames and check results to the main model to fix. Done when: Every code check and yes/no question passes, the mix sits on its targets, and the judge agrees with human labels at high-80s or 90s on the golden set. Proof: @shannholmberg (37,522 followers), 2,525 saves / 1,200 likes / 76,793 views, 2026-10-19 https://x.com/shannholmberg/status/2112104872451756348; @akshay_pachaar (292,113 followers), 193 saves / 189 likes / 18,144 views, 2026-10-27 https://x.com/akshay_pachaar/status/2115163583764169031; @undefinedKi (20,049 followers), 2,748 saves / 1,249 likes / 210,749 views, 2026-08-03 https://x.com/undefinedKi/status/2084279627204235703; @hanakoxbt (24,233 followers), 2,205 saves / 1,469 likes / 152,358 views, 2026-07-21 https://x.com/hanakoxbt/status/2079692306337153079; varg how to skill "Check a render against its reference by Refs" https://varg.ai/how-to/skills/refs-check-render; varg how to skill "Motion energy check by Scenario" https://varg.ai/how-to/skills/scenario-motion-energy-check; varg how to skill "OpenMontage Sound Design Rules" https://varg.ai/how-to/skills/openmontage-sound-design; varg how to skill "Shot QA" https://varg.ai/how-to/skills/craft-shot-qa ### 09.3 Before delivery, run a slop-tell pass over frames, voice and script, list every tell and where it is, and fix each one. Without a reference the model falls back to centered text, a gradient background and everything fading in, so many videos look the same. Viewers spot over-smooth skin in a second, check the product most closely, and called a synthetic product shot fake. Do: - Give the model 1-2 reference videos and check the render does not fall back to centered text, gradient background and everything fading in. - Check motion for flat fades, stiff robotic motion without micro-movements, unrequested slow motion, smooth camera moves and broken physics. - Check people for over-smooth skin, perfect light, morphing or duplicated characters, and hands for extra fingers. - Replace any frame where the product drifts with the real packshot, and check it for new text, logos or watermarks. - Have a second listener check dialogue for gibberish and for delivery that is too theatrical, too monotone or like a read-through. - Cut script tells (binary contrasts, throat-clearing openers, fake-profound endings, short punchy sentences), then compare with the last 30 published videos and kill anything that rhymes. Done when: The audit list of slop signs is empty, no product frame drifts from the real packshot, and nothing rhymes with the last 30 published videos. ```text "Take the AI look out of this page. Don't touch features, content or section order. Only change the craft. 1. Audit: list every sign of AI slop and where it is. Start with purple gradients, glows, gradient text, emoji labels, "New ✨" badges, all-caps labels with wide letter spacing, animations that never stop, spacing that's bigger here and smaller there. ``` Proof: @rexan_wong (24,918 followers), 18,188 saves / 6,888 likes / 647,682 views, 2026-09-26 https://x.com/rexan_wong/status/2103707054108299437; @Mnilax (10,079 followers), 1,631 saves / 722 likes / 115,719 views, 2026-10-11 https://x.com/Mnilax/status/2109314840057565345; @FynCas (45,181 followers), 104 saves / 109 likes / 7,236 views, 2026-10-28 https://x.com/FynCas/status/2115525477272097131; varg how to skill "How to prompt Kling 4 for cinematic AI video" https://varg.ai/how-to/skills/kling-4-prompting; @CharlieDFilm (4,171 followers), 6,063 saves / 6,040 likes / 1,271,505 views, 2026-09-13 https://x.com/CharlieDFilm/status/2098929162735997071; @gokayfem (7,262 followers), 110 saves / 183 likes / 382,361 views, 2026-10-26 https://x.com/gokayfem/status/2114819508032684289; @DavidRunsAds (3,517 followers), 1,235 saves / 661 likes / 52,674 views, 2026-09-27 https://x.com/DavidRunsAds/status/2104301098227888180; @Ciri_ai (11,908 followers), 31 saves / 233 likes / 30,457 views, 2026-11-02 https://x.com/Ciri_ai/status/2117143506443784677; varg how to prompt "30s breakdance montage from three pose stills" https://varg.ai/how-to/prompts/yt-kling4-breakdance-montage; varg how to prompt "Coca-Cola's World Cup film, AI can shots called fake" https://varg.ai/how-to/prompts/yt-rtquwigalus; @HindeJ72752 (10,542 followers), 82 saves / 85 likes / 6,787 views, 2026-11-01 https://x.com/HindeJ72752/status/2116701457974235326; @second_bassoon (2,953 followers), 3,399 saves / 2,474 likes / 303,110 views, 2026-11-02 https://x.com/second_bassoon/status/2117307653320007954; @second_bassoon (2,953 followers), 3,399 saves / 2,474 likes / 303,110 views, 2026-11-02 https://x.com/second_bassoon/status/2117365915385839995; @petergyang (298,515 followers), 13,099 saves / 6,265 likes / 1,198,614 views, 2026-07-22 https://x.com/petergyang/status/2079943830024188105; @deedydas (261,658 followers), 4,381 saves / 2,099 likes / 159,803 views, 2026-09-29 https://x.com/deedydas/status/2104957026199900220; @ScottyBeamIO (78,454 followers), 1,684 saves / 674 likes / 125,929 views, 2026-09-19 https://x.com/ScottyBeamIO/status/2101280444914258251; @Voxyz_ai (27,685 followers), 241 saves / 118 likes / 10,068 views, 2026-10-31 https://x.com/Voxyz_ai/status/2116515537496993856 Top 10: 1. Without a reference the model falls back to its default look: centered text, gradient background and everything fading in; check for it. (https://x.com/rexan_wong/status/2103707054108299437) 2. Make the agent render, check and tweak in a loop, auditing the output frame by frame, until the result looks right. (https://x.com/charliejhills/status/2116936969251164475) 3. Have someone other than the maker listen for AI gibberish in dialogue; the maker's own ears missed it. (https://x.com/second_bassoon/status/2117365915385839995) 4. Compare generated speech with the locked voice track; Seedance still alters speech slightly, speeding or slowing it with the speaker's movement. (https://x.com/ZssBecker/status/2113497202702115226) 5. Layer review: programmatic checks first, an LLM judge second, and humans last, used only to calibrate the judge. (https://x.com/undefinedKi/status/2084279627204235703) 6. Split review: code checks measurable rules, a decision model inspects frames, and the main model only reviews flagged issues and fixes the code. (https://x.com/shannholmberg/status/2112104872451756348) 7. Give review to its own subagent that watches the render second by second and lists problems by severity without touching anything. (https://x.com/Voxyz_ai/status/2104194355556671987) 8. Run an independent critic subagent that compares screenshots of the work against top reference work, and do not let the builder stop until the critic says it matches. (https://x.com/anshuc/status/2081801979131818412) 9. Score the render against the reference film on cut timing, brightness, motion, loudness and format, then make frame-exact fixes. (https://varg.ai/how-to/skills/refs-check-render) 10. Measure whether the video parks: percent of still frames, longest hold, motion and cuts per 10 seconds, plus loudness gates. (https://varg.ai/how-to/skills/scenario-motion-energy-check) ## 10. GitHub Save every approved step, so a bad take is one command away from undone. ### 10.1 Commit each approved step as a small checkpoint, and rewind to the last good one instead of patching a bad result. HiTw93 builds in the smallest usable steps with commits so work can roll back anytime, and says commits need your OK. damkina7 warns against agents editing many files without atomic checkpoints: rewind instead of patching a bad patch, squash only on full green. Do: - Break the job into the smallest usable steps and commit after each one. - Work in plan mode (claude --permission-mode plan) so edits wait for approval, and get the person's OK before every commit, release or deletion. - Check the diff after each task; a small request can come back as dozens of changed files with test assertions deleted. - On a failed step, hard-reset to the last checkpoint instead of patching the bad patch. - Squash a step's commits only when every check is green. - Add a PreToolUse hook that blocks force-push, .env writes and rm -rf before they run. Done when: Every approved step sits in its own commit that the person OKed, and any later step can be undone by resetting to the previous checkpoint. ```text Add this to ~/.claude/settings.json. Keep every hook, permission and plugin I already have. Then create the two agent files below and show me the diff. ``` Proof: @HiTw93 (165,160 followers), 1,030 saves / 735 likes / 46,075 views, 2026-10-29 https://x.com/HiTw93/status/2115805352520282623; @damkina7 (828 followers), 128 saves / 95 likes / 7,365 views, 2026-11-02 https://x.com/damkina7/status/2117241955415400550; @0xarcos (365 followers), 24 saves / 37 likes / 1,099 views, 2026-11-02 https://x.com/0xarcos/status/2117245427254722840; @LimestoneHQ (13,239 followers), 125 saves / 58 likes / 18,835 views, 2026-10-30 https://x.com/LimestoneHQ/status/2116082213016273313; @PatchOnChain (2,029 followers), 1 saves / 14 likes / 972 views, 2026-11-01 https://x.com/PatchOnChain/status/2116995297561952610 ### 10.2 Keep each video as diffable text: one .md per stage plus composition and render command, approved style in brand/, large media out of git blobs. li9292: text video source can go into git, be diffed and forked. HeyGen keeps composition, storyboard, assets and render command together; beamnxw's harness is one .md per stage. A CLAUDE.md rule saves approved style to brand/. Crab's author warns that large files stored as ordinary blobs make repos slow. Do: - Add a CLAUDE.md rule: "save every brand color, font, logo and approved style to brand/". - Keep one .md file per stage: brief, style, storyboard, motion, sound and render. - Keep each video as an editable composition next to its storyboard, assets and render command. - In the workflow text, mark the fixed skeleton, the swappable slots and the frame-exact time anchors. - Commit only compact pointers for large media and write the files to an object-storage bucket, never as ordinary Git blobs. - Save hand edits into the film's source files so the agent keeps them on its next edit. Done when: The video re-renders from text in the repo: stage .md files, composition, storyboard and render command sit together, brand/ holds every approved style element, and large media are pointers, not blobs. ```text save every brand color, font, logo and approved style to brand/ ``` Proof: @li9292 (19,879 followers), 2,391 saves / 1,179 likes / 109,214 views, 2026-09-15 https://x.com/li9292/status/2099687200975720699; varg how to skill "HeyGen's Real Launch Videos as HyperFrames Source" https://varg.ai/how-to/skills/heygen-hyperframes-launch-videos-source; varg how to skill "Opus 5.5 Motion Studio Harness" https://varg.ai/how-to/skills/opus-5-5-motion-studio-harness; @thegreatest_sv (10,183 followers), 1,134 saves / 583 likes / 47,172 views, 2026-10-20 https://x.com/thegreatest_sv/status/2112609040962359354; @haipingfu (917 followers), 208 saves / 223 likes / 15,823 views, 2026-09-01 https://x.com/haipingfu/status/2094863027682386215; @Asteri_eth (3,340 followers), 52 saves / 68 likes / 3,329 views, 2026-10-29 https://x.com/Asteri_eth/status/2115758546067349910; @IamEmily2050 (49,235 followers), 68 saves / 51 likes / 4,297 views, 2026-11-03 https://x.com/IamEmily2050/status/2117438054478786620 ### 10.3 After each good result, write the lessons into a skill, save approved looks as commands, and test the skill against a no-skill run. ashebytes updates skills at the end of each process until they one-shot; anshuc had Claude dump its learnings into a skill on GitHub. The prayag_sonar card turns every good clip into a /command. Voxyz says many installed skills make no difference, so test with and without. Do: - At the end of each video, update the skills with what produced a decent outcome. - Save each approved look as a command file that takes the subject and a one-line brief as inputs. - Keep each skill as a folder with SKILL.md, human docs, references and template assets, and a precise one-line description, which is all the agent sees until it loads the skill. - Run one short request in two fresh subagents, one reading the skill and one not, and delete the skill if the results cannot be told apart. - Package skills, subagents, hooks and .mcp.json as one plugin directory so one command rebuilds the setup. Done when: The with-skill result differs visibly from the no-skill run on the same request, and the skills install on a new machine with one command. ```text Test whether actually does anything: 1. Come up with one short request this skill should handle, like "build a pricing page" or "write a login endpoint". 2. Spin up two fresh subagents and give them nothing but that request. A reads the skill first, then does it. B must not use the skill or read its files. 3. Show me both results side by side and list every difference. If you can't tell them apart, tell me the skill is dead weight and I can delete it, then stop. If you're testing all of them, end with a list of which ones to delete. 4. If it's a skill I wrote, write down whatever A still gets wrong, edit the skill file directly, and rerun from step 2 with the same request until you can't find anything wrong ``` Proof: @ashebytes (45,332 followers), 973 saves / 591 likes / 44,357 views, 2026-10-27 https://x.com/ashebytes/status/2115215620279263329; @anshuc (17,494 followers), 1,246 saves / 686 likes / 57,634 views, 2026-07-27 https://x.com/anshuc/status/2081801979131818412; varg how to prompt "Slash-command motion styles reel built in Claude" https://varg.ai/how-to/prompts/x-2116545733197541787; @Voxyz_ai (27,685 followers), 115 saves / 101 likes / 7,488 views, 2026-11-02 https://x.com/Voxyz_ai/status/2117372427714089436; @0xchewa (4,457 followers), 8,538 saves / 3,247 likes / 598,877 views, 2026-07-12 https://x.com/0xchewa/status/2076107957277008100; @DanKornas (100,742 followers), 68 saves / 56 likes / 3,483 views, 2026-10-24 https://x.com/DanKornas/status/2114076072946102644; @Scenario_gg (23,292 followers), 1,658 saves / 1,047 likes / 59,912 views, 2026-11-02 https://x.com/Scenario_gg/status/2117286675693060430; @thedelost (5,536 followers), 24 saves / 38 likes / 1,259 views, 2026-11-02 https://x.com/thedelost/status/2117361587975492093 Top 10: 1. Save every approved look as a reusable slash command so the style library grows. (https://varg.ai/how-to/prompts/x-2116545733197541787) 2. Store each approved piece with a note on why you changed it, so the agent has something specific to apply. (https://x.com/shannholmberg/status/2098850243760808268) 3. Keep the video source as text so it can go into git, be diffed, and be forked with a few slots swapped. (https://x.com/li9292/status/2099687200975720699) 4. Add a CLAUDE.md rule that saves every brand color, font, logo and approved style to brand/. (https://x.com/thegreatest_sv/status/2112609040962359354) 5. Have the agent build in the smallest usable steps and commit each one so you can roll back at any time. (https://x.com/HiTw93/status/2115805352520282623) 6. Update the skills at the end of each video once you reach a decent outcome. (https://x.com/ashebytes/status/2115215620279263329) 7. Store each video as an editable composition kept together with its storyboard, assets and render command. (https://varg.ai/how-to/skills/heygen-hyperframes-launch-videos-source) 8. When a step goes wrong, rewind to the last good checkpoint instead of letting the agent patch the bad patch. (https://x.com/damkina7/status/2117241955415400550) 9. Test each skill you install or write against a run without it on the same request. (https://x.com/Voxyz_ai/status/2117372427714089436) 10. Keep compact pointers in git and write large media as chunked, deduplicated files to an object-storage bucket, as Crab does without an LFS server. (https://x.com/haipingfu/status/2094863027682386215)