Step 1: Find references
Without a reference, the agent falls back to its default look: "centered text, gradient background, everything fading in." That's @rexan_wong after going through the Claude motion videos everyone was sharing (18K bookmarks). It's why so many AI videos look alike.
"Most AI videos look cheap because nobody gives the AI a real film reference." (@ScaleWthAI)
So don't describe the look. Show the agent 1–2 real frames and tell it to match them.
Film stills: you browse, the agent studies them
- Film-Grab: full sets of stills from one film, browse by title. Free. The post that recommended it got 5.8K bookmarks.
- ShotDeck: film frames you filter by color, light, camera angle and location, with HEX color search. Shot cards list palette, lens, lighting and film stock. Paid, from $13 a month.
- Frameset: film stills; @EXM7777 pairs it with ShotDeck.
- Shot Cafe: reference shots for a specific kind of scene.
- Eyecannndy: the names of visual techniques, so you can ask for them by name.
- Pinterest: boards other people already curated: costumes, color grades, posters.
Sources the agent searches itself
- Refs: 4,322 launch films, demos and motion pieces, measured shot by shot and searchable from your agent over MCP. Start here for product videos.
- FrameThrower: stills from 5,489 films, searchable by lighting, lens, shot size, color and mood.
For web and app UI: "pick 3 to 5 sites you like on Godly, Refero or One Page Love, drag the screenshots straight into Claude Code, and have it tell you what makes each one look good" (@Voxyz_ai, 1.4K bookmarks).
How to hand references over
- Match, don't describe. Attach the frames and ask the agent to match them.
- Give every reference one job. Say what it sets and what it must not copy: "The reference image defines WHO the person is, not the pose, location, expression, clothing, lighting, or composition" (@swiat_ai, 1.5K bookmarks).
- Write down what each one lends. One line per frame in refs.md: palette, framing, light, texture or motion.
- Pull clips, not only stills. "yt-dlp → pull reference clips and audio from the web" (@flxrnc, 2.7K bookmarks).
- Borrow the look, not the content. Palette, framing, lens and light. Never actors, logos or recognizable people. Keep links, not copied frames.
Prompts to paste
No references yet? We wrote this whole step as a skill, find-references: the agent goes through seven sources in order and comes back with style, film, pop-culture and trend references and ten visual ideas on one board. Here it is as one prompt. In Claude Code, paste it after /goal and the agent keeps working until every line under "Done when" is true. Our test run on a sample brief: 11 minutes, 20 references, 10 ideas, nothing paid.
/goal Find references and 10 visual ideas for this video: [what it is, who it's for, platform and aspect ratio, length, brand, 1–2 things I love, 1 thing I hate].
Go through these 7 sources in order and save what you open to refs/:
1. varg how to (curl -s "https://varg.ai/how-to/api/search?q=<format and look>"): style and prompt cards that already worked for this format. Search results carry no numbers: open the best 2–3 cards (add .md to the link) and copy their installs, stars or bookmarks.
2. X, last 7 days: posts in this format or look. Filter with min_faves:, then rank by bookmarks yourself. The prompt is in the post, its link or the author's first reply. Write likes, bookmarks, views and date into trends.md. No X access? Use web search and say so.
3. Refs (refs.video/mcp): launch films, demos and ads like this one. Free search gives title, length, summary and thumbnail; the shot-by-shot data needs a Refs account.
4. FrameThrower (framethrower.ai/api/mcp): film frames by light, lens, shot size, color and mood, plus its ads and animation libraries.
5. Film-Grab, Shot Cafe, Eyecannndy, ShotDeck: they block bots, so never fetch them. Write search links for me into links.md: film-grab.com/?s=<query>, shot.cafe/tag/<tag>, eyecannndy.com/technique/<technique>, shotdeck.com (search needs a login).
6. Pinterest and Are.na: search by attribute, not subject ("<film> still", "<decade> film grain", "<character> cosplay"). Download Pinterest images with gallery-dl; for Are.na, search it and put the channels in links.md.
7. YouTube: music videos, trailers, title sequences, game cinematics and ads people remember. Pull a 5–15 s clip with yt-dlp only where motion or pacing matters.
Find four kinds of reference: style (palette, light, texture, type), film (framing, lens, camera move), pop culture (moments people recognize) and trends (what people saved this week).
Rules: open every reference you keep and never describe a frame you didn't see. Borrow the look, never actors, characters, logos or recognizable people. Found frames are for words and judging, not for feeding an image or video model. Skip the first, obvious pick (for a sci-fi look, not Blade Runner). Copy every number exactly as the source shows it.
Done when:
- refs.md has 12+ references you opened: 3+ style, 3+ film, 2+ pop culture, 3+ trends from the last 7 days with their numbers. Each has its link, what it lends (palette, framing, light, lens, texture, motion), one job and what it must not copy.
- links.md has the searches only I can open.
- ideas.md has 10 visual ideas. Each combines 2+ references of different kinds and has a look line, the first shot, why now (a trend with its numbers, a different one per idea where you can; wild ideas may skip it) and the risk. 3 are close to this week's trends, 4 are stretches, 3 are wild. None falls back to centered text, a gradient background and everything fading in.
- board.html shows every reference and idea, numbered, with sources, so I can pick.
- You listed every source you could not reach.
Already have the reference? Adapted from the brief @higgsfield_ai published with a launch film made with zero video generations (522 bookmarks):
Use the attached [reference] video as the primary reference. Inspect its compositions, typography, scene changes, motion, and audio before building. Ask only for information that cannot be recovered from the reference.
Let the agent do it
- find-references: this step as a skill. Seven sources, four kinds of reference, ten visual ideas, one board. Install: npx skills add vargHQ/skills --skill find-references
- media-download: downloads a Pinterest search or board, and clips from X, YouTube, TikTok and Instagram, into refs/.
- look-references: this whole step as a file the agent reads, with a refs.md template.
- varg-mockups: puts every reference on one board, so you approve each frame before anything is generated.
- Launch video from a reference: the full route on varg how to, from a Refs film to a checked render.