agentleFS
Maksim-BurtsevMaksim-Burtsev/ai-film-crew/skills/scene-set/SKILL.md

scene-set

Skill0 starsChanged 18 days agoMITView on GitHub

Lock the geometry of a multi-shot AI video in one set file before any picture, then check every generated still against it - room, furniture, props, every person's seat and facing, one camera per setup. Use when planning shots or stills, writing image prompts for a scene, checking a still before paying for video, or when shots of one scene disagree (a moved table, swapped seats, flipped eyelines, wrong heights, light from the wrong side).

The file

We split this skill into the rules it gives an agent

2 rules. About 0 are common, 2 are distinctive, and 0 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈2

    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
  1. Distinctive:

    Every shot is a view of one set.

    line 8
  2. Distinctive:

    Write where each person's mic, mug, notes and glasses are, which hand holds what, and the mount of every mic.

    line 43

≈ “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.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Set first

Every shot is a view of one set. People, furniture and light never move between shots; only the camera does. Stills prompted one by one, without a set, drifted: they turned the table, moved people across the room, stood up someone who should sit and flipped eyelines between the wide and the close-ups.

Two scripts, run from the project root with uv run:

  • scripts/scene.py: the set file in, the plan, prompt cards, layout images and fact questions out. Its docstring documents the set file (TOML, metres, top view) and the optional cast file.
  • scripts/frame_check.py: a generated still or take in, one PASS/FAIL line per check out.
  1. Write set.toml from a real reference. Walls, window, light side, the table and its axis, every prop, each person's place, posture, facing and garment colours, one camera per setup (talk = true on cameras that carry lines). Start from examples/ (a podcast desk, a round podcast table, a couch, a meeting table). For a video that already exists, read the set off its widest approved frame, then check it against every approved shot.
  2. scene.py plan set.toml plan.png: the top view, cameras, the 180° line. It fails a camera across the line or a face into the lens. Show this picture with the storyboard: people read a room from it.
  3. Prompts come from the set. Each still prompt opens with scene.py card set.toml <cam> verbatim and passes scene.py block set.toml <cam> block.png as the layout image (text alone never held a layout). scene.py refs adds each person's cast images and the sentence that numbers them.
  4. Check every still before it becomes video. frame_check.py set.toml <cam> still.png compares the faces in the still with what this camera should see: how many, which side, who is closer, whose head is higher, which way each looks, which side the light comes from, how big the speaker's face is. A FAIL stops the still. Re-roll it or fix the set; never buy a take from it.
  5. Ask the facts too. scene.py ask writes fact questions for one frame ("which side is the window on?"); a blind reviewer answers them (see blind-review) and scene.py grade compares. Vision models are poor at "are these shots consistent?" and fair at "which side is the window?": the reviewer answers facts, the set decides. scene.py sheet puts plan, layout, candidate and neighbours in one image.

Done when: every planned shot names its camera, plan prints no warnings, and every approved still passes frame_check.py and grade (or each FAIL was looked at and answered).

What frame_check can't see

It measures faces, not meaning. It misses a medium shot drawn where the set wants a close-up, and a hand over a face reads as no face. A FLAG means "look yourself". When it disagrees with your eye, the blind review decides.

Props are part of the set

Write where each person's mic, mug, notes and glasses are, which hand holds what, and the mount of every mic. A prop's side can turn the model's eyeline: a mic on the outer side turned a speaker away from the host in 7 of 11 stills. Make a person's second still an edit of the first so mounts and props match. Give the blind review the prop questions (how many mics, which side, what the hands hold).

Coverage that holds up

  • One close-up per speaker, all from one side of the line. Lines only in close-ups, never in a wide.
  • A silent wide only for an action the story needs (≤ 1.5 s, nobody facing the lens).
  • Two people the set puts far apart share a frame only in a wide.

Download now (ZIP)

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
What did you use it for? e.g. Reviewing pull requests in a Rails app

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.