recurring-cast
Skill0 starsChanged 18 days agoMITView on GitHub
Keep a recurring AI character the same person across shots and videos (face, age, body, props, voice), and check it with a face model calibrated on your own judged stills. Use before the first paid still of a person, when a face drifts or a lookalike appears, when someone says "that's not her", or when picking a face, gesture or voice for a character.
In this repository4 other files it gives its agents.
- blind-reviewSkill
- product-sizeSkill
- scene-setSkill
- sound-mixSkill
The file
We split this skill into the rules it gives an agent
4 rules. About 0 are common, 4 are distinctive, and 0 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈4
Distinctive · estimate
This author's own choices. Worth a look.
0
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 17
Real photos, or a still the client approved: Never an earlier video's stray frame, which compounds drift.
- Distinctive: line 48
All stills of one person side by side: before any lip-sync, on the face and on the background.
- Distinctive: line 53
One quality tier per video: every face gets the same model, resolution and framing floor.
- Distinctive: line 56
Check at full size, both sides of the head.
≈ “Common” is an estimate: at least 1 in 5 of the 372,429 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Recurring cast
A viewer forgives a lot, but not a face that changes. Each failure below once reached a client.
One card per person, before the first paid still
The card holds: the face's source, the look (hair, glasses, garments, jewellery, which wrist), height and build, posture and gestures (in and out), the voice and its direction. A taste the owner or client states in chat ("he takes up space", "never a plain tee") goes into the card the same day. One unwritten taste once cost two hours of gesture rounds.
Where a face comes from
- Real photos, or a still the client approved. Never an earlier video's stray frame, which compounds drift. At most one edit pass away from that source.
- A turnaround sheet for a person who appears from many angles: six head angles (front, both 3/4, both profiles, back) plus full length front, side and back, with height. Pass the view nearest the camera's angle, the real photo (it decides face and age), and the front; never the sheet itself (a multi-view sheet gets copied as twins). Image models read about four character images at most.
- Mark flawed views (one reads older, one has the wrong hair) and never pass them.
- Age only at the client's own number, one pass. Never "younger" or "fresher" on your own.
Check the face with a model, calibrated on your stills
scripts/identity.py turns each face into numbers with a face model and compares them with the person's real
photos (5 to 20, in one folder).
- Calibrate once per person:
uv run scripts/identity.py calibrate --refs photos/ana --same approved/ --other rejected/. It scores stills you judged "the same person" and "someone else", prints both ranges and a threshold in the gap, or says this model can't tell them apart on your stills. - Score every still:
identity.py score stills/*.png --refs photos/ana --min <threshold>. - Score every take against its own first still:
identity.py score take.mp4 --against first_still.png --min 0.80. Video and lip-sync models drift a face mid-take; 0.80 held on lip-sync clips, re-check it on yours.
A score screens. It never picks the winner and never excuses a reject: the side-by-side look and the blind review decide. Faces under ~250 px tall, profiles and covered eyes score like strangers.
Which face model. SFace, the default (Apache 2.0), catches a different person. On one cast it scored the other
characters 0.10-0.27 against 0.86 for the person herself, but it scored a lookalike actor 0.52-0.55, where ArcFace
gave 0.38-0.41. ArcFace (InsightFace w600k_r50) separates lookalikes better, but its published weights are licensed
for non-commercial research only: for client work, license it or use a commercial face-comparison service, and pass
--model arcface --arcface <path> only with a copy you're allowed to use. Whatever the model, calibrate on your own
stills: thresholds don't carry over between models.
More checks
- All stills of one person side by side before any lip-sync, on the face and on the background. Each pose can pass alone while the final shows three slightly different people.
- Face size: a face under ~250 px tall in a 1080x1920 frame doesn't hold. Video models morph it, and image models draw a different person. Redraw a small head on a 4x crop from the refs and paste it back. A close-up is its own still, never an upscaled crop.
- One quality tier per video: every face gets the same model, resolution and framing floor. A visible difference costs more than the dollar it saves.
- Props on people are identical in every shot: glasses whole (both rims, bridge, temples), the same jewellery. Check at full size, both sides of the head; contact sheets hide these.
Keep the face through an image edit
Edit models redraw what they weren't asked to: an edit that adds glasses narrows the eyes and puffs the lids, and the
person reads as someone else. scripts/keepface.py registers the edit onto the base still by the face and takes the
face back from the base, pixel for pixel:
keepface.py glasses base.png edit.png out.png: the base's face with only the edit's new glasses frame on it.keepface.py face base.png edit.png out.png: the edit everywhere (a new pose, hands, a prop) except the base's face. It fails when the two layers disagree at the seam (two mics, a smeared collar).keepface.py check base.png out.png: do the eyes, nose and mouth still line up with the base?
New poses are edits of the approved still, never fresh generations. The thresholds were measured on one person's edits (black frames on light skin); the glasses mask looks for a dark frame, so check light or rimless frames at full size.
Taste on raw candidates
A choice of face, gesture or voice goes to the owner as cheap raw candidates, labelled raw. Checks and polish run only on the pick: one pose was polished through four reviews, then dropped in two minutes.
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.

