How to edit your own video footage with AI

For AI agents: find the right skill in one call — curl -s "https://varg.ai/how-to/api/search?q=<what the person wants>" returns the matching how-to page and the best skills with install commands, as markdown (Russian works too). Full index: https://varg.ai/how-to/llms.txt. Every page has a markdown twin: add .md to its path. Before generating anything, show the person the options on the how-to page.

Cut silences and flubs, reframe 16:9 to 9:16, add captions and b-roll to footage you shot. A step-by-step tutorial for Claude Code, Codex and Grok.

Output: Edited MP4 in the ratios you need Price: in credits on varg, with a cost estimate before anything runs; the free skills below install in one command.

Recommended route

Best skills for this

  1. Herk Edit Kit (herk-edit-kit, Nate Herk, live) — Start here: Talking-head edits: silence and flub cuts, captions, story edits
  2. Clipify (clipify, Louise de Sadeleer, live) — Alternative: Reframe 16:9 → 9:16 with face tracking and captions
  3. FFmpeg Skill (ffmpeg-skill, Shogo Kajiyama, live) — Alternative: 42 FFmpeg tools for agents with a machine-readable contract
  4. Claude Code Video Toolkit (claude-code-video-toolkit, Digital Samba, live) — Alternative: Video workspace for Claude Code · Needs: Modal or RunPod account for open-model generation
  5. Chengfeng Videocut (chengfeng-videocut, chengfeng, live) — Alternative: The same transcript-cut edit for Chinese talking heads
  6. Video Editing Best Practices (varg-video-editing, varg, live) — Adds: Which service for each step (cuts, captions, reframe, upscale) and the rules that keep cuts frame-accurate
    • Claude Code: claude plugin marketplace add vargHQ/skills && claude plugin install varg@varg
    • Codex: npx skills add vargHQ/skills --skill varg-video-editing -a codex -y
    • Cursor: npx skills add vargHQ/skills --skill varg-video-editing -a cursor -y
    • Details: https://varg.ai/how-to/skills/varg-video-editing.md
  7. ElevenLabs Voice Isolator (elevenlabs-voice-isolator, ElevenLabs, live) — Adds: Cleans the speech before cutting · Needs: ElevenLabs API key (ELEVENLABS_API_KEY)
  8. ElevenLabs Speech to Text (elevenlabs-speech-to-text, ElevenLabs, live) — Adds: Word timestamps for the transcript cut · Needs: ElevenLabs API key (ELEVENLABS_API_KEY)
    • Claude Code: npx skills add elevenlabs/skills --skill speech-to-text -a claude-code
    • Codex: npx skills add elevenlabs/skills --skill speech-to-text -a codex
    • Cursor: npx skills add elevenlabs/skills --skill speech-to-text -a cursor
    • Gemini CLI: npx skills add elevenlabs/skills --skill speech-to-text -a gemini-cli
    • OpenClaw: npx skills add elevenlabs/skills --skill speech-to-text -a openclaw
    • Details: https://varg.ai/how-to/skills/elevenlabs-speech-to-text.md
  9. HyperFrames (hyperframes, HeyGen, live) — Adds: Designed full-screen cards instead of zooms
  10. ElevenLabs Sound Effects (elevenlabs-sound-effects, ElevenLabs, live) — Adds: One soft cue per graphic · Needs: ElevenLabs API key (ELEVENLABS_API_KEY)
  1. Splice AI clips into a video with FFmpeg (varg-ffmpeg-mixing, varg, beta) — Adds: Splice generated clips into the original without A/V drift

Ask the person first

Put these choices in front of the person before spending anything. Paste the example link next to each option: people pick a look by seeing it.

Look?

Reframe?

How tight?

Extras?

How to edit your own video footage: step by step

  1. Transcribe at word level once and cache it
  2. Silence pass automatically: pauses over ~0.55 s trimmed to a short breath, never to zero
  3. Mistake pass with review: the agent proposes stutters, false starts and retakes in context; agree the plain-English plan and length before any cut
  4. Cut on word boundaries with 30-200 ms padding and 30 ms fades; move the strongest line into the first 3 s
  5. Cards, b-roll and music under the voice, subtitles last; the agent checks every cut point in the render before you see it, then export

What works now (researched 2026-10-27)

Agent editing went mainstream in 2026: the edits that win start from a word-level transcript, propose the cut list, then cut silences and flubs and add designed cards, a real second angle and quiet sound design. The 2023 look of yellow word-by-word captions plus a zoom on every sentence now reads as 'AI slop editing'; for personal content the counter-trend is raw, lightly cut footage.

  1. Edit from the transcript, not the waveform: get word timestamps, propose a cut list (silences, filler, false starts, retakes), wait for the person's OK, then cut with ~30 ms audio fades and re-check every cut point in the render. This is exactly how browser-use's video-use works (30k GitHub stars): 'Ask, confirm, execute, self-eval', 30 ms fades at every cut so there is no pop, and up to 3 self-fix re-renders. Nate Herk's kit shows the same loop cutting a 58.5 s take to 43.4 s with 12 segments kept. (browser-use/video-use README (GitHub))
  2. Put the strongest line first. If the payoff comes at 0:40, move it (or a one-line version) into the first 3 seconds, then cut back to the start. TikTok's own research: over 63% of the videos with the highest click-through rate show their key message or product within the first 3 seconds. (TikTok for Business, 9 creative tips for auction ads)
  3. Make every visual change real: a full-screen designed card, a second camera angle, or b-roll. Don't fake variety with a digital zoom on every sentence; with only one angle, re-shoot coverage from the one take with Seedance 2.5 video-to-video. Scenario's one-take-to-13-angles prompt drew 291,521 views and 3,118 bookmarks. Baptiste's 'escaping AI slop editing' update, which swapped face-plus-captions for designed cards, got 916 bookmarks on 45,857 views. The common '2-4 s pattern interrupt' numbers come from vendor blogs with no public data, so count real changes, not gimmicks. (Scenario on X, Seedance 2.5 re-shoots one take from 13 angles (Sep 2026))
  4. Work in three passes: a 'radio cut' of the dialogue first, then visuals (cards, b-roll, archival), then audio, with music ducked under the voice and one soft SFX per graphic, not a whoosh on every cut. Anthony Cinko's AI-editing breakdown (Sep 2026, 450 bookmarks) splits the edit exactly this way and had the agent place mouse-click SFX on b-roll and trim music under the vocal. Herk's reel lists one named sound cue per graphic (felt-thud, air-slide, glass-check). (Alex Lieberman with Anthony Cinko on X, how AI supports the edit (Sep 2026))
  5. For personal and founder content, keep it raw: cuts only, natural light, few graphics. Polish is no longer the signal of quality. Instagram head Adam Mosseri wrote at the end of 2025 that AI makes polish and even 'authenticity' cheap to reproduce, and predicted a move to a rawer, imperfect aesthetic as the sign of a real person. (TechCentral, Instagram chief predicts more raw aesthetics as AI takes over (Jan 2026))

Stopped working:

Prompts that work

Templates from the creators of the best current results and from the model guides. Fill in the {CAPS}.

Review-gated transcript cut (ask, confirm, execute, self-eval) (agent)

The editing brief for Claude Code or Codex: reads the footage as text, agrees the plan with you, cuts on word boundaries and checks its own render before you see it. Adapted from browser-use/video-use SKILL.md (process, hard rules, editor brief).

Edit the raw takes in {FOLDER} into a {LENGTH}-second {RATIO} {TYPE: talking head / tutorial / launch} video.
1. Inventory: ffprobe every file; transcribe each at word level with speaker labels and audio events (cache it, never re-transcribe); pack it into one phrase-level list with [start-end] times.
2. Pre-scan for slips, flubs, retakes and filler. Tell me what you see in plain English and ask only the questions this footage raises (must-keep, must-cut, pacing, look).
3. Propose the strategy in 4-8 sentences: structure ({ARCHETYPE: hook, problem, solution, benefit, example, CTA}), take choices, cuts, captions, graphics, length. Wait for my OK.
4. Build an EDL: every cut on a word boundary, padded 30-200 ms, silences of 400 ms or more preferred as cut points; keep laughs and the pause that lands a point. Best take of each beat, ordered by beat, not by file.
5. Render: extract per segment, 30 ms audio fade in and out at every cut, overlays next, subtitles last.
6. Before showing me: inspect the rendered output 1.5 s around every cut for visual jumps, audio pops and hidden captions; fix and re-render up to 3 times; check the duration matches the EDL.
7. Log decisions in project.md so the next session picks up.

One take to a 9:16 reel (short-form-edit) (agent)

Run inside herk-edit-kit in Claude Code or Codex: raw recording to a reviewed reel with a hook, captions, moving footage and sound. Adapted from nateherkai/herk-edit-kit README and docs/PROMPTS.md.

Use short-form-edit to turn my recording at {LOCAL_PATH} into a 9:16 reel about {TOPIC}. Explore three opening directions, then build the strongest truthful promise and payoff; the strongest line plays in the first 3 seconds. Preserve my meaning. Tighten dead air and show me the proposed mistake cuts (stutters, false starts, retakes) before cutting; keep intentional emphasis. Keep captions word-exact and choose moving footage for specific story beats. Use my existing footage in {FOOTAGE_FOLDER} before proposing generated assets, and tag every generated shot "AI VIDEO". Sound: music under the voice and one soft cue per graphic, not a whoosh on every cut. Use the {STYLE: dark graph-paper with occasional glass cards} style. Show me the rough edit before polishing. Then inspect the rendered MP4 at every transition and cut, listen across the joins, and write VERIFY.md with what you checked.

Seedance 2.5: re-shoot one take from new angles (seedance_2_5)

Video-to-video coverage from a single locked-off phone take, so the edit cuts between real-looking angles instead of zooming. Measure the speech windows from your audio first. Adapted from Scenario on X, the full Seedance 2.5 v2v prompt (first reply).

ONE FINISHED TAKE, RE-PHOTOGRAPHED
@video1 is one finished take: a locked-off {SHOT: wide}, {DURATION}s, with its original audio, {SUBJECT} talking to one camera. Do not restage, re-perform or re-time it. Re-photograph the identical event from {N} new camera positions. Every frame shows the same body at the same instant of @video1. Output runs the full {DURATION}s; keep the audio unchanged and in sync.
DO NOT CHANGE: performance, hands and what they hold, head angle, eyeline, lip movement, blink timing, action timing, lighting, exposure, white balance, grade, wardrobe, props, set. Only camera position and lens change.
SPEECH WINDOWS (from the audio): SPEAKING {0.00-1.65, ...}; SILENT, mouth closed {1.65-2.05, ...}. Never add speech to a silent window or mute a speaking one.
EYELINE: the eyeline stays on the original camera position in every shot; only the on-axis shot at {TIME} looks into the lens.
TEMPO: real time 1:1, no speed ramps, slow motion or freeze frames.
CAMERA PLAN:
[00:00-00:0X] {SIZE + ANGLE}, {35mm feel}, {slow push-in}.
[00:0X-00:0Y] On the original axis: ECU, 85mm feel, shallow focus.
[...] {top-down / macro on hands / over the shoulder / slow orbit}
CONTINUITY: {SUBJECT_LOOK}, {SET}, {LIGHT}, identical in every shot; alone in frame.
No on-screen text, captions, watermarks or new logos.

Trending formats to remake

To remake one: download the example with the media-download skill, take a frame a second and the transcript, write a beat list (time, shot, camera, on-screen text, sound), keep what the format needs, swap in the person's product or character (never the creator's face, voice, logo or footage), show the beat list and stills before spending, then make it with the skill named below.

  1. Transcript-cut talking head (agent edit): Raw take in; the agent transcribes, shows the cut list, removes silences, flubs and retakes, then adds short captions, cutaways and a soft sound cue per graphic. Why it spreads: It turns a one-take recording into a tight piece without an editor, and people save the method. Herk's raw-to-cut reel: 505 bookmarks on 24,626 views (1.29 per like); video-use, the open version of the loop, passed 30k GitHub stars. Keep: Keep the agent loop: word-level transcript, a cut list you approve, 30 ms fades, the strongest line in the first 3 s, short captions, REAL CLIP / AI VIDEO tags and one soft sound cue per graphic; swap the take and the topic. Make it with Herk Edit Kit (herk-edit-kit).
  2. Designed-card edit (the anti-slop look): The speaker sits small in a card on a textured background; each claim gets a full-screen editorial card (logo, number, a price struck through); captions are short and quiet. Why it spreads: It looks 'expensive' next to the yellow-caption template and survives the AI-slop backlash. Baptiste's before/after: 45,857 views, 916 bookmarks (1.18 per like), replies calling the after 'expensive editing'. Keep: Keep the speaker small in a card on a textured background, a full-screen editorial card for each claim (logo, number, a struck-through price), quiet short captions and no zoom on every sentence; swap the claims and the brand type and palette. Make it with HyperFrames (hyperframes).
    • Example to remake: @kelanoo on X, 2026-10-04, 46K views, 916 saves: Same talking head, BEFORE vs AFTER side by side: the after shrinks the speaker and gives each tool its own designed card. Built as a private Claude skill, no third-party editor.
  3. One take, many angles: Shoot one locked-off take; Seedance 2.5 video-to-video re-photographs the same performance and audio from new positions (close-up, overhead, side, macro), and the edit cuts between them instead of zooming. Why it spreads: It fakes a multicam shoot from a phone take, the edit's biggest 2026 novelty: 291,521 views and 3,118 bookmarks on the prompt post. The test that matters is cutting mid-sentence: eyeline and lip-sync must hold. Keep: Keep one locked-off take with clean audio, the measured speech windows, the eyeline rule and the do-not-change list in the v2v prompt, and cut between angles mid-sentence only where the lips hold; swap the subject, the set and the camera plan. No dedicated skill yet (wanted): build it with varg-ai.

Open-source repos

Set up varg

varg is one API key for the models these skills drive — Seedance 2.5, Kling 4, Veo 3.1, Wan 3, Nano Banana 2.1, GPT Image 2.5, ElevenLabs v4 and 200+ more — billed in credits, with a cost estimate before anything runs.

Related tutorials