ai-video
Skill1 starsChanged 14 days agoMITView on GitHub
Plans and prompts AI-generated video shots (Higgsfield, Seedance, Kling, Wan, Veo) that have to sit inside a real edit — a mascot or presenter that stays the same character in every shot, a person in existing footage swapped for another character, a short cinematic hook in front of real screen recordings. Covers character sheets, start frames, video prompts, lip-sync and how to spend as few credits as possible. Use it whenever someone wants to add or replace shots with AI video, keep an AI character consistent, turn a mascot into a moving presenter, swap a person out of a video they already recorded, or asks how to prompt Higgsfield or a similar video model.
3 other projects use a version of this.
Adopted by
Most recently taken up by
- 0xadvait/ai-video-skilltheir own version, first seen Nov 8, 2026
- TimoteoAltobrando/ai-video-skilltheir own version, first seen Oct 15, 2026
- tylerprogramming/claude-skillstheir own version, first seen Aug 9, 2026
Other versions3 projects use a version of this.
Word for word · 0
None found.
Edited · 0
None found.
Wrote their own version · 3
- 0xadvait/ai-video-skill★ 9same name
- TimoteoAltobrando/ai-video-skill★ 4same name
- tylerprogramming/claude-skills★ 0same name
How matches are found. Word for word: the exact same text. Edited: shares passages with this file (“99% same” means 99% of its five-word phrases also appear here; 10% or more counts). Their own version: similar meaning, or, weakest, the same name. We can’t tell which project had it first. Projects are counted among the 200 most-starred candidates, and mentions are links to the repository we found, up to 50 per site, so both counts are a floor.
In this repository6 other files it gives its agents.
- AGENTS.mdAGENTS.md
- clip-cutterSkill
- content-directorSkill
- footage-editSkill
- reference-researchSkill
- vertical-videoSkill
The file
We split this skill into the rules it gives an agent
2 rules. 0 are common, 0 are distinctive, and 2 are very specific to this project.
0
Common
Most of the 3 edited versions we read keep these.
0
Distinctive
This author's own choices. Worth a look.
2
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Very specific: line 67
Test at 720p or with the cheaper model tier, change one thing per attempt, watch the whole clip (drift shows up in the middle), and only …
- Very specific: line 76
Keep the original narration timing from narration.mp3.
“Common” is measured: kept by at least 6 in 10 of the 3 edited versions of this file we could read. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
AI video shots inside a real edit
AI video is the expensive, unreliable part of the job. Everything here exists to spend credits only on shots that are already decided, and to make the character come back the same every time.
The kit itself does not call any AI video service. It prepares what the generator needs, writes the prompts, and takes the results back into the edit (vertical-video or footage-edit). The person runs the generations in their own account.
Decide what is really generated
Go through the existing video shot by shot and put each one in one of four boxes. Most of a how-to video stays real.
| Box | When | Tool |
| Keep | Screen recordings, UI, anything the viewer must see exactly as it is | the edit |
| Swap | A real shot is right but the person in it must become someone else | Genjutsu-style character swap, see references/swap-real-footage.md |
| Insert | Only a region changes: a face inside a phone camera box, a logo on a screen | generate the subject alone, composite it into the region |
| Generate | No footage exists: an establishing shot, a hook, a closing beat | image → video, see references/prompting.md |
Swap before generate. A swap keeps the real camera move, timing and cuts, so it lines up with the narration that is already there; a generated shot has to be cut to fit afterwards.
The flow
1. Cut the shots that will change
python scripts/prep_shots.py video.mov --detect
python scripts/prep_shots.py video.mov --out shots/ --shot intro=0-2.9 --shot desk=2.9-7.89--detect lists where the picture changes, so the times come from the footage.
Each shot comes out as an H.264 clip plus its first and middle frame, with the
full soundtrack and a manifest.json. Use the best original you can get: a
WhatsApp copy is about 464 px wide and every AI tool will show it.
2. Lock every character before any video
Still images are cheap or unlimited; video is not. Build each character's
sheet first and get it approved — references/character-sheets.md:
- Turnaround (front, 3/4, side, back) in one image on mid-grey.
- One clean face close-up: this is the anchor every video prompt points at.
- Expressions (including mouth shapes for lip-sync) and the poses the script needs.
Save each character as a reusable reference in the tool (Higgsfield calls them
Elements, @name).
3. Make every start frame as an image
For Generate shots, build the first frame of each shot as a still, with the approved sheet as reference, and fix it there. A wrong start frame only gets worse once it moves.
4. Then video, cheaply first
Test at 720p or with the cheaper model tier, change one thing per attempt,
watch the whole clip (drift shows up in the middle), and only render the
chosen version at full resolution. Prompts: references/prompting.md.
Which model for what: references/models.md.
5. Back into the edit
Generated clips go in like any other footage: public/media/ in a
vertical-video project, or the raw/ folder of a footage-edit cut.
Keep the original narration timing from narration.mp3; if the voice is new,
generate it to the same script and let voice.py re-time the scenes.
Rules
Never put the real screen in the generator. Interfaces, numbers, QR codes and button labels come back warped. Keep those shots real and change only what sits around or inside them.
One character description, pasted verbatim into every prompt. Rewording it between shots is how a character drifts.
Mid-grey behind characters, never white. White melts into light clothes and leaks into the subject; grey keeps edges clean for compositing.
One face per reference image when generating video. Sheets with many faces make video models switch between them.
Rights. Swap only footage you own or are allowed to use, and only into characters you have the rights to. A real person's face needs their permission.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.