agentleFS
PicsArtPicsArt/gen-ai-skills/skills/film-cast-design/SKILL.md

film-cast-design

Skill7 starsChanged 21 days agoMITView on GitHub

Make a film's people, props and places consistent on pictures - line-keyed descriptors, one sheet per person and one plate per prop compiled by the server, reading an asset review, cutting each panel into its own view, locking the approved asset, and the user's own photos kept as they came. Use when a film needs its cast, props or places designed, pictured, locked or repaired.

The file

Show its rules and text

We split this skill into the rules it gives an agent

15 rules. About 0 are common, 8 are distinctive, and 7 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈8

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 7

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Every person and every prop the story handles gets its pictures.

    line 27
  2. Distinctive:

    When the tone is an era, generated people are dressed and groomed to it.

    line 83
  3. Very specific:

    Use picsart-video-rules's image default unless the user chose another model.

    line 91
  4. Very specific:

    Keep the selected look.camera: references retain its image texture, colour and detail.

    line 97
  5. Very specific:

    Do not substitute a portrait-camera look or add unrequested lens and exposure settings.

    line 99
  6. Very specific:

    Keep look.scheme and look.hue in it: left off, the cast comes back in no colour while the shots compile to the film's hue.

    line 100
  7. Very specific:

    Never type or retype a template.

    line 105
  8. Distinctive:

    Every independent picture goes in the same batch.

    line 113
  9. Distinctive:

    Every tile comes back approved, changed or discarded.

    line 117
  10. Distinctive:

    No separate still test is bought.

    line 138
  11. Very specific:

    Before any video, for every scene the run spans: every tag it names has a locked row (picsart_asset_list with the project folder, read …

    line 149
  12. Distinctive:

    Read the places in story order: after a fire, later scenes need frames from the run that burned the room, not its clean pictures.

    line 152
  13. Distinctive:

    If the view is sound, the take is repaired (film-shot-generation).

    line 161
  14. Distinctive:

    Two faces blending in a two-shot: only each person's identity and wardrobe views, a medium two-shot rather than a wide, and never a …

    line 163
  15. Very specific:

    When a source changes, check only what depends on it, by the job it did (identity, wardrobe, geometry, light) and record reuse, recrop, …

    line 167

≈ “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

Film cast

The model has no memory. What keeps a face the same across forty shots is a locked picture of it, cut into single views, passed to every video that shows that person. Read these before the step:

FileRead it before you…
references/views.mdcut a view out of a sheet, or check one
references/own-photos.mduse a photo the user brought of a person, a product or a place

What each asset gets

AssetPicturesCompiled with
A person you generateone sheet, the strip of Picsart's canvas: full-body front, side and back and a large face close-up in one row, in one generationpicsart_film_compile_prompt, kind: "sheet"
A person the user photographedthe photo itself; a profile and a sheet only for views the photo lackskind: "profile", then kind: "sheet", from the photo
A propone plate: the object alone on a neutral ground, and the object beside an everyday thing for scale; no people, no handskind: "plate"
A placea descriptor now; its pictures come from its scene (film-shot-generation)—
A crowdone plate with the dress code, the age spread and what they carry; each shot states the exact countkind: "plate"

Every person and every prop the story handles gets its pictures; nothing is tiered. A figure in the background, or an object nobody touches, gets no asset: the shot that shows it describes it.

A state (wet, bloodied, coat off) is a panel on the person's sheet with its own tag (@cal_wet), generated from the approved sheet so it stays the same person. A change that lasts most of the film may earn a sheet of its own; that is a craft call. A place at night, or after a fire, is a scene with its own pictures, never a state.

The starting point

Draft every descriptor first, as below. Then OFFER the cast board once: picsart_asset_choose_characters with the film's title as film and the whole cast (at most 12 a board), each character's descriptor lines as story. Every card opens on the ready-made characters; a card left with nothing picked keeps those lines, so a board sent untouched changes nothing. It is an offer, not a gate: it spends nothing, and a reply that ignores it keeps every character as the story has them.

Its answer, character_chooser_feedback, has one entry per @tag:

  • story — keep your descriptor.
  • preset with saved — reuse that saved asset as the user’s sheet; compile no sheet and spend nothing.
  • preset without saved — it has no sheet yet. Include its credit cost in the quote, compile with the answer’s sheetModel, then save it under the preset’s own name so a later pick reuses it.
  • preset without saved or person — the answer carries the look as reference: pass it unchanged as that character's reference on picsart_film_compile_prompt. A person is a character the user built on the board: a person, and any of their features, face and hair, outfit, accessories and body, already joined into that one reference. It holds who they are, how they look (an animal's species and coat, the face, the build) and what they wear and hold. Add from your descriptor only the paths fromStory names: identity.<field> for what the picks leave open — a ready-made character names no origin, and a person names no hair unless a face and hair look was picked — and outfit when no clothes were picked. Accessories picked without clothes are already in faceAndOutfit: keep them there. A field you add replaces the preset's. Presets carry no voice: the voice line stays yours. The descriptor you save is that reference in your own line keys, plus the voice.
  • describe — rewrite the descriptor from the user's words. Keep your voice line and every line the words leave unsaid.

A person the user photographed is not on the board: the photo is the asset. Then carry on: one plan, every sheet together.

Descriptors

Draft every descriptor in one turn, from the story, the cards and the references; mark what you invented as a design proposal. They are line-keyed (face, hair, jacket, shoes, …), so a change is one line, and they are concrete: no evaluative words.

  • A person: sex, build, face shape, eyes, hair and facial hair; costume item by item with material and condition; shoes; accessories; posture, gait, hand habits, tempo. Age goes in the compiler's identity fields, never as a number in prose. The base costume is what most scenes show; scene gear is a state.
  • A place: architecture, materials, key objects, condition, atmosphere, one motivated light per time of day, the look's colour, and one memorable anchor object. With a colour scheme, the look's colour is the film's palette line, verbatim; with the scheme on Auto, each place names its own colours.
  • A prop: shape, material, condition, distinguishing details, and size relative to a known object, never a bare number.
  • A voice line for everyone who speaks, before their first spoken shot: accent, register, timbre, habitual cadence. Chosen from the character and frozen once saved; film-shot-generation' Voice says how the prompts carry it.
  • When the tone is an era, generated people are dressed and groomed to it. If the story's period and the tone's era disagree, flag it with both options and restyle nothing until it is settled.

The descriptor is the sheet's prompt and the asset's record. In a video the views carry identity instead, as film-shot-generation says.

Pictures

  1. Use picsart-video-rules's image default unless the user chose another model. Record it once per film. Check its image inputs and supported generation ratios with picsart_model_params; frame/crop a supported ratio to the film canvas instead of silently changing models.
  2. Compile every picture with picsart_film_compile_prompt and its reference kind, passing the film's look, the same object the shots get, and the descriptor's lines in reference (identity or faceAndOutfit, outfit, a plate's descriptor and anchor, assetPalette). Keep the selected look.camera: references retain its image texture, colour and detail; neutral identity views omit an action camera's fisheye and a drone's aerial viewpoint. Do not substitute a portrait-camera look or add unrequested lens and exposure settings. Keep look.scheme and look.hue in it: left off, the cast comes back in no colour while the shots compile to the film's hue. Pass the film's image model as model: a picture sent with reference images then gets a ratio that model takes with them. It returns the prompt, the negativePrompt, the aspectRatio and a promptToken. Send the prompt as returned, and only parameters the chosen model supports; omit negativePrompt when unsupported. Never type or retype a template. With a supplied photo, set reference.usePhoto: true and put its original URL in the actual imageUrls; the flag alone attaches nothing.
  3. Compile with verifyWith (request with the imageUrls you will send, film with the saved assets, their identityReferences and views, limits from picsart_model_params) and read each result's check: picsart_film_verify_prompt already run on that prompt. A failure is fixed in the inputs and compiled again; a skipped reference budget or view check has not run. Never retype a compiled prompt into verify.
  4. Sheets are generated at the model's highest resolution, so every panel has pixels to spare.
  5. Every independent picture goes in the same batch. A failed one blocks only what depends on it.

Reading the review

Every tile comes back approved, changed or discarded.

  • A change is surgical: the named line or panel, nothing else. A point defect on an otherwise sound sheet is a masked edit with an image edit model, not a re-roll.
  • Approved is locked: cut the views the cards need (references/views.md) from the approved full-resolution sheet, look at each, and save with picsart_asset_save: tag, assetType, name, descriptor, sheetVersion ("v1"; a new version is a new save, never an overwrite), voice (an object: mode preset or design, or nonSpeaking: true; voiceId, model, description, language, accent, sampleUrl), portraitUrl (the sheet, or the user's photo), the views as referenceImages (the close-up first, as references/views.md orders them), identityReferences with the user's original photo URLs (or [] when every view is generated), the scenes it serves, projectFolderUid, status: "locked". The result's locked, not its saved, says it is done. A locked save lands in the project's "Library", anything else under its "Assets"; with no project folder it goes to the flat global "Plugin Assets" folder. Only a save registers an asset for picsart_asset_list. A draft or review save returns locked: false with a remaining list. stressTest is recorded only, never a lock condition: pass it when the user asked for a spot-check and it ran; a battery with failures blocks the lock until fixed or noted in waived.
  • Discarded siblings are never used.

Check the cast on the scene pictures

No separate still test is bought. A face locked on one good picture can drift on real shots, and the scene pictures (film-shot-generation) are where the cast first appears in its rooms, light and company: check every face you can read there against its locked views, as you would a take, and beside the people they share a frame with, since someone who holds up alone often breaks in a two-shot. A person who drifts has an under-written descriptor: tighten the lines the picture got wrong, record them on the row, and they go into every later prompt. Say what you saw in the pictures' review; a new sheet or picture is generated only when the user asks for the change, quoted on its own.

The gate into a run

Before any video, for every scene the run spans: every tag it names has a locked row (picsart_asset_list with the project folder, read back), and every tag that needs a panel, a state or a profile resolves to a single-view file. A photo-only row cannot answer a view it lacks. Read the places in story order: after a fire, later scenes need frames from the run that burned the room, not its clean pictures.

Repairs after a run

  • A drifting face: check the view first. A wrong crop, a neighbour's edge in the box, or a sheet sent whole explains most drift, and the fix is a re-cut, not a generation.
  • Repair from the locked views, never from the last take. A take made from a take inherits its drift and adds its own.
  • If the view is sound, the take is repaired (film-shot-generation). After the same face drifts twice, tighten its face lines and re-roll. A sheet whose face is too small is regenerated and re-cut.
  • Two faces blending in a two-shot: only each person's identity and wardrobe views, a medium two-shot rather than a wide, and never a composite of two solo pictures.
  • A descriptor changed after takes exist does not silently regenerate them. Name the takes it affects and flag them for review.
  • When a source changes, check only what depends on it, by the job it did (identity, wardrobe, geometry, light) and record reuse, recrop, repair or rebuild. A version bump alone does not cascade.

Efficiency

  • A reference picture is scoped to its job: a neutral ground, no location, no scene light. The compiler's asset style line carries the look without the staging; never paste the scenic one on.
  • Fix a contradiction in a prompt (a neutral ground and a location; a tight face and a full body) by changing the clause, never by stacking negatives.
  • Record a failure's defect separately from its likely cause. Two failures of the same kind mean the input or the method changes; never send a third identical attempt.

Traps

  • A sheet or a user's composite sent whole to a video model, with the panel named in words. No pointer works: cut the panel.
  • Cutting before approval, or from a preview instead of the full picture.
  • Guessing the grid instead of reading the boxes from the pixels.
  • A sheet on a scenic background; two people on one sheet; faces too small to cut.
  • Saying an asset is locked because the save came back saved.
  • Asking the user to approve their own photo, or retouching it in place.

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.