agentleFS
PicsArtPicsArt/gen-ai-skills/skills/movie-producer/SKILL.md

movie-producer

Skill7 starsChanged 21 days agoMITView on GitHub

Make a short narrative film from a story, with recurring characters, one look and approved cast/location references, then draft video and a reviewed cut. Storyboard or animatic only on explicit request. Use for short films, movies, multi-scene scripts, or continuing an existing film (shot list, shot changes or next step). Load first; it brings in the film skills for each stage. Not for a drama series or its episodes (make-short-drama), a single clip, or footage the user already has.

The file

Show its rules and text

We split this skill into the rules it gives an agent

25 rules. About 0 are common, 16 are distinctive, and 9 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈16

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 9

    Very specific

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

  • Common
  • Distinctive
  • Very specific
  1. Very specific:

    When you read picsart_scene_capabilities or the recipe index (picsart_scene_recipe without name), pass useCase: "movie-producer": the …

    line 16
  2. Distinctive:

    Read picsart-video-rules first.

    line 25
  3. Very specific:

    Before acting on a stage, load the skills in its row that apply to this job, also when the work resumes there.

    line 28
  4. Distinctive:

    Read these before you need them:

    line 45
  5. Very specific:

    The challenger at each decision: Before you go on from a decision, quality-review's challenger (the challenger agent where the host has it) …

    line 107
  6. Distinctive:

    Read the user's brief first: Every file in the material that is not media, read in full before the first question, as picsart-video-rules's …

    line 116
  7. Distinctive:

    Ask for the idea only if the user has not given one.

    line 120
  8. Distinctive:

    If the genre changes at any point, film-story-writing rewrites the story first.

    line 140
  9. Very specific:

    Read each one's picsart_model_params before the board (film-shot-generation's model steps 2 and 3).

    line 157
  10. Distinctive:

    Check each boundary: can the next shot fit in the preceding run within the model's duration and reference limits?

    line 190
  11. Distinctive:

    Write the internal cut pairs and each necessary run edge using film-shot-generation's seams reference: match action, eyelines and screen …

    line 198
  12. Distinctive:

    After a long stretch on one video, offer once to reopen the board to recap what comes next.

    line 212
  13. Very specific:

    Look at each picture (picsart-video-rules, Asset URLs) before anything uses it, then put them all on picsart_asset_approve in one call (at …

    line 238
  14. Distinctive:

    Treat "Change one", "not her", "profile is not ok", "it's not her", or any no to a cast picture as a rejection.

    line 253
  15. Very specific:

    Before regenerating, call picsart_asset_save for the rejected version with status: "rejected", its original tag, sheetVersion, portraitUrl, …

    line 257
  16. Very specific:

    Never show a rejected picture again, including in a comparison, previousUrl, an old plan card or an asset recap.

    line 267
  17. Very specific:

    Check rejectedAssets[] before every display, reference selection, plan and resumed film turn.

    line 271
  18. Distinctive:

    Remove rejected references from pending opening frames and shot plans before continuing.

    line 272
  19. Distinctive:

    Keep the location map. Follow film-shot-generation's scene-pictures reference: a wide for every scene, its opening scene picture cut from …

    line 280
  20. Distinctive:

    Do not ask them to choose the shots again.

    line 318
  21. Distinctive:

    Before any run is quoted, a shot that film-shot-generation flags for a palette arguing with the film's scheme gets one short ASK (with your …

    line 329
  22. Distinctive:

    When the last take lands, film-cutting assembles the film once: every candidate take in story order with hard cuts, every generation one …

    line 353
  23. Distinctive:

    When the takes look glossy or plastic, as generated footage often does, one option is the realism grade, applied as the …

    line 379
  24. Distinctive:

    A dropped call is not a failed generation: follow its job as film-shot-generation says.

    line 403
  25. Very specific:

    An outage is status, not a decision: Keep going on everything free, keep the blocked item in openDecisions, and try again next turn.

    line 408

≈ “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

Short film

When you read picsart_scene_capabilities or the recipe index (picsart_scene_recipe without name), pass useCase: "movie-producer": the answer leads with what this job reads and names the rest in one line.

A film is not a long video. The model has no memory between generations: describe a character loosely and he arrives in every shot with a different face, jacket and age. Everything here exists to keep those people and places consistent. Lock the story, shot list, cast and location references, then shoot draft video directly. The first moving draft the user reviews is the assembled cut. Storyboard panels and an animatic are made only when the user explicitly requests them. They are not part of the default flow, and are never offered as another decision before shooting.

Read picsart-video-rules first. One job: a story in, a short film out, with recurring people and places and one look.

Before acting on a stage, load the skills in its row that apply to this job, also when the work resumes there. (preview-requested) applies only to an explicit request for a storyboard or animatic; (animatic-requested) only to an explicit animatic request. These skills and the preview REVIEW are otherwise skipped; the location references are still prepared.

StageGateSkills
1. The idea——
2. The lookASKfilm-story-writing
3. The story and the modelASKfilm-story-writing, film-shot-generation, sd2-pe, sd25-pe
4. The shot listASKfilm-story-writing, film-shot-generation
5. Cast, places and propsSPEND, ASK, REVIEW (preview-requested)film-cast-design, film-shot-generation, scene-building-basics, film-animatic (preview-requested), voice-over-generation (animatic-requested), quality-review (preview-requested)
6. ShootingSPEND, ASKfilm-shot-generation
7. The cutSPEND, ASK, REVIEWfilm-cutting, scene-building-basics, quality-review
8. FinishingSPEND, ASK, REVIEWfilm-finishing, background-music, realistic-color-grade
9. DeliverASK, DELIVERfilm-finishing, footage-reframing

Read these before you need them:

FileRead it before you…
references/stages.mdsay a stage is done, or open the next one
references/board-replies.mdact on a reply from any board
references/project-record.mdwrite or resume the project record

Before you start

Route first. The dividing line is recurring identity across scenes.

The user bringsRoute
One clip's worth of idea ("a boot stepping onto scree")Not this use case. Say so, and offer one generation: picsart_model_choose, then SPEND (quote it in one line, then get the yes with your host's own question tool, such as AskUserQuestion, "Yes, go ahead" / "Yes, up to N credits for this task" / "No", never as a typed question; only where the host has no such tool, those options numbered in one message; unless a spend limit the user stated covers it: N credits or an explicit no-limit ("as much as needed"), never inferred from words like "go ahead". Within it, ask nothing: quote it in one line ("About 8 credits; 12 of 20 left after this"), call it, then say what it spent and what remains, keeping a running total that includes failed calls, and stop to ask before the first call that would cross it. A money amount is asked once to restate it in credits; in unattended pace, never ask: call only with a consent or a stated spend limit already recorded, otherwise write the quote to pendingSpends and stop, as picsart-video-rules says) and picsart_generate
Footage to cut togetherNot this use case. Say it is not built yet here
A story: scenes, people who recur, dialogueThis use case, from stage 1
A film already under way (a project record)Resume: read the record, say where the film stands, continue at its stage

Pace: director by default, as picsart-video-rules says. In express pace the user comes in only for the boards, the three locks (the look, the story with its model, the assets), the cut's review, every charged call no stated spend limit covers and anything beyond a quote.

How you talk to the user

Talk and ask as picsart-video-rules says. Here that means:

  • one board per decision; everything else waits in the record's openDecisions;
  • what each board's reply holds is in references/board-replies.md;
  • where things stand reads like "Stage 5 of 9 · cast, places and props — locking Cal (2 of 5 done)", and the gate view shows which scenes have everything they need ("not ready yet", never "a hole");
  • an alarm word is for a real gate violation, such as a video about to be bought for a scene whose assets are not locked.

The project record

The film's state is one record, and it outranks your memory of it. Where the host gives you files, keep it as film.json in the working folder; otherwise keep it in the conversation and give it to the user once, at the end of a working session, to paste back next time. Never print it mid-flow. Update it whole at every milestone: a gate passed, an asset locked, a batch dispatched or landed, a review round, picture lock, the export. Its shape and the migration of older ten-stage records are in references/project-record.md.

Beside the facts it keeps the brief: the logline, the world in three sentences, the people as people, and the standing decisions with their reasons ("hard cuts only: the film is about abruptness"). On resume, read it before saying where things stand. With no record at hand, take what the user says about the film as its state and carry on; ask for the record only when the step needs a fact they have not given.

Gates

A gate is passed only when its checklist in references/stages.md is met, or when the user consciously overrides it; record the override and its reason. Before you say a stage is done or name the next one, write its checklist with ✓ or ✗ per line and the evidence beside each ✓ (a status a tool returned, a user's yes, a recorded override). A line with ✗ means the stage is not done: say which line and what closes it.

Choosing how to meet a rule is your craft call; choosing to spend or risk money against a closed gate is never yours. Which model shortlist, which order, whether a face needs one more view: yours, recorded, never asked. Buying video for a scene whose assets are not locked, going beyond a quote, trading quality for money: the user's.

The challenger at each decision. Before you go on from a decision, quality-review's challenger (the challenger agent where the host has it) asks "why this?" of it: the idea (stage 1), the look and the story (2, 3), the shot list (4), each sheet, plate, scene picture and location map picked and why (5), and every quote before its SPEND (5, 6, 8, and a repair in 7 that needs a generation). Act on its verdict as quality-review says. A stop, or a rethink of the user's own choice, goes to the user in the next ASK, the cheaper alternative first, before anything is spent.

1. The idea

  • Read the user's brief first. Every file in the material that is not media, read in full before the first question, as picsart-video-rules's The user's own brief says. A brief sets the message, the audience and the must-shows.

Ask for the idea only if the user has not given one. Once there is one, ask nothing more here: runtime and frame are set on the look console, and the story proposes an ending the user can change.

2. The look

picsart_film_setup comes first, before a word of story is written, one call per visual world. Genre decides how the story is told, and every shot card sits under what this console locks: the style line, the lens family the cards' focal lengths come from, and the colour scheme each card's own palette has to fit.

  • Pre-set it from the idea with suggested, one suggested.why line each, as film-story-writing says: the genre the events make, a tone when the story names a period, and the colour.
  • film (the title) is required. ratio and duration go inside suggested.selections, under those field names, never as top-level arguments: the schema rejects anything it does not list.
  • Pass ratio whenever the user named a format at all ("vertical", "TikTok", "square"). Pass the runtime they said as duration. Nothing infers a format; the console's default stands otherwise.
  • The console is the ASK. On film_setup_feedback (verdict: "locked"), store stylePrefix verbatim and the selections as they came. That lock is the look's lock; there is no later one.
  • If the genre changes at any point, film-story-writing rewrites the story first.

3. The story and the model

The story. film-story-writing writes it, for the locked genre, with its viewer track. State the identity budget you inferred ("two faces and one garage") rather than asking for it. Show the track's after-watch questions under the story, for correction: they are what the user and the cold watch are asked after each draft.

The model, in the same turn. The candidates are the current models, in this order:

ModelFrameLength per runReference imagesDraft
seedance-2.516:9, 9:16, 1:1, 4:3, 3:4, 21:94–30 s30draft: true
kling-v4-omni16:9, 9:16, 1:1, 21:93–30 s8none: 720p
wan-3.0-video16:9, 9:16, 1:1, 4:3, 3:42–30 s10none: 480P
veo-3.116:9, 9:164, 6 or 8 s3none: 720p

Read each one's picsart_model_params before the board (film-shot-generation's model steps 2 and 3); the table is a guide, the params decide. Drop a candidate only when its params cannot hold the film's frame or its runs (Veo 3.1 holds no 1:1 and no run over 8 s). Never offer an older model (Seedance 2.0, its Fast and Mini, Kling V3, Hailuo, Wan 2.7) unless the user names it. The user's balance and the cheapest-first rule set the draft resolution, never the model: they never put an older model on the board or make one recommended.

Open picsart_model_choose under the story with the candidates that survive, in that order, the first one candidates[].recommended, each candidates[].bestFor line written as this film's trade ("one continuous 30 s take, tops out at 720p").

  • The pick is the story's yes unless the user sends a note with it. A note about what happens is answered as film-story-writing says, and the rewritten paragraph is shown again with the model board.
  • On the pick, record the model, its take ceiling and its edit and extend siblings, as film-shot-generation finds them. One model for the whole film. A Seedance pick (seedance-…) also brings in the sd25-pe skill for Seedance 2.5 or sd2-pe for Seedance 2.0, as film-shot-generation's Seedance reference says.

4. The shot list

picsart_film_shotlist once, with every scene in one call. Load film-story-writing before you write the cards: it turns the approved story into them, and its peak rule marks one card, the Peak line's shot, with the moved face unguarded and what shows ("@teo relief, unguarded: the tears come"), every other face held in; film-shot-generation packs them into runs against the model's ceiling, with one join sentence on the shot that opens each run. Pass film, the model, and the setup's defaults, so the user never picks from a vocabulary the film has not adopted.

Use the model's capacity before opening the board. Load film-shot-generation's Runs method and its seams reference. Read the selected model's accepted durations and reference limits from picsart_model_params, then pack the ordered shot cards into the fewest runs that fit, across scene boundaries. Assign run ids only after packing: the board preserves those boundaries, so one id per beat would lock in needless generations. Several cuts belong inside one generation; a climax, a close-up, a new scene or "easier to generate" is no reason to split it.

Check each boundary: can the next shot fit in the preceding run within the model's duration and reference limits? If yes, keep packing; if no, record the limiting constraint. Preserve shot order, shot lengths and dialogue; do not stretch shots or add filler to reach the ceiling. The final remainder can be short. A shorter run otherwise needs a specific model limit or the user's explicit split. For example, six proposed runs of 7, 7, 4, 3, 6 and 3 seconds become one 30-second generation when the selected model accepts 30 seconds and their combined references fit. The nine shot cards and their internal cuts stay separate inside that generation.

Write the internal cut pairs and each necessary run edge using film-shot-generation's seams reference: match action, eyelines and screen direction, keep geography, and carry frames only for its named exceptions. Fewer generations must preserve this method, not replace the cuts with one long shot or buy a bridge for every boundary.

  • The board approves camera and seconds. It is not a place to rewrite the story: an edit that changes what happens is a story note, so rewrite the paragraph first, then open the board again.
  • On shotlist_board_feedback, replace your list with it, runs included. Fields the board does not carry (references, prompt versions, trims) stay on the shot by id.
  • A long scene held in one shot comes with film-shot-generation' coverage recommendation on its card. Keep the single shot only if the user chooses it on the board.
  • A card that film-shot-generation flags for a kinetic move on a big face, set on the board or changed later, gets one short ASK (with your host's own question tool, such as AskUserQuestion, so the options are buttons the user taps, the recommended one first; only where the host has none, the options numbered in one message; in unattended pace, never ask: write it to the project record's openDecisions with the answer you assumed, as picsart-video-rules says) before the run is compiled: the move as the user set it, or one of film-shot-generation' safer versions. The camera stays the user's.
  • After a long stretch on one video, offer once to reopen the board to recap what comes next. Honour a no without asking again.

5. Cast, places and props

The project folder first, before anything is generated. picsart_drive_create_folder with path: "Picsart Plugin/<film-slug>" and children: ["Library", "Assets", "Shots", "Exports"]. Keep every uid it returns in the record. Say it once, plainly: "Created Picsart Plugin/the-garage in your Picsart Drive, with Library, Assets, Shots and Exports inside. Everything this film makes is filed there." Never say it again, and never list the whole Drive: no list without a folderUid. Pass the right folderUid on every picsart_generate and picsart_job_status call of the film, and the project folder as projectFolderUid on every asset call. A finished job is filed only when you save it, as picsart-video-rules's SPEND rule 4 says: save every finished picture and take into its folder before you show it. Never call a filed file permanent.

Then film-cast-design's conveyor, in one pass:

  1. Draft every descriptor as film-cast-design says, then offer the cast board once (film-cast-design, The starting point). It is the one stop here and costs nothing; a reply that ignores it keeps the story's descriptors. There is no other approval round.
  2. Compile each picture with picsart_film_compile_prompt (a reference kind) and read its check, then one SPEND for the whole batch: every sheet and plate on one quote.
  3. Dispatch on the yes as one batch (picsart-video-rules, SPEND), with folderUid set to Assets, and let the batch's card follow the jobs as picsart-video-rules says.
  4. Look at each picture (picsart-video-rules, Asset URLs) before anything uses it, then put them all on picsart_asset_approve in one call (at most 12 tiles; more go on a second board), with the one thing to judge them by as the board's judge, and one line per tile as its assets[].why.
  5. Act on the reply as references/board-replies.md says. Each approved asset is cut into views and locked with picsart_asset_save, as film-cast-design says. Before regenerating a rejected picture, save that version as rejected and record it as below. The replacement is regenerated under its own SPEND and shown alone on a fresh board, with assets[].version; omit assets[].previousUrl when the previous picture was rejected. A discarded tile is recorded as rejected and never used.

A person the user photographed is saved as it came and is not put up for approval: it is theirs. film-cast-design flags what it cannot settle on its own, about a photo the user brought or about the film's era. Raise each flag once, as one short ASK (with your host's own question tool, such as AskUserQuestion, so the options are buttons the user taps, the recommended one first; only where the host has none, the options numbered in one message; in unattended pace, never ask: write it to the project record's openDecisions with the answer you assumed, as picsart-video-rules says) with the options film-cast-design names. Anything generated is its own SPEND.

A rejected cast picture

Treat "Change one", "not her", "profile is not ok", "it's not her", or any no to a cast picture as a rejection. If several pictures were shown and the target is unclear, ASK (with your host's own question tool, such as AskUserQuestion, so the options are buttons the user taps, the recommended one first; only where the host has none, the options numbered in one message; in unattended pace, never ask: write it to the project record's openDecisions with the answer you assumed, as picsart-video-rules says) which one without showing them again; do not regenerate until it is identified.

Before regenerating, call picsart_asset_save for the rejected version with status: "rejected", its original tag, sheetVersion, portraitUrl, projectFolderUid and saved asset metadata. Pass the saved identityReferences too. Every new save supplies the user's original photo URLs there, or identityReferences: [] when all views are generated. Do not infer provenance from a view's position; preserve unknown provenance on older records. Record its tag, version, file uid, URL and all extracted view/panel URLs in the project's rejectedAssets[], and mark that version's asset row rejected. Confirm the save succeeded before a replacement is generated; on failure keep the local rejected entry and retry the save, never reuse it. A replacement gets a new version and stays in review until the user approves it.

Never show a rejected picture again, including in a comparison, previousUrl, an old plan card or an asset recap. Never pass any part of it as a reference: no full-body front, profile, crop, view or panel. Rejecting one view rejects the whole sheet unless the user explicitly said otherwise; record any such exception and keep only the explicitly permitted view separate from the rejected sheet. Check rejectedAssets[] before every display, reference selection, plan and resumed film turn. Remove rejected references from pending opening frames and shot plans before continuing.

For a person made from the user's own photos, keep those original photos in identityReferences[] and use them as the identity reference for every replacement. Rejecting a generated sheet never rejects the user's original photos; do not substitute a rejected generated face for them.

Location references before shooting

Keep the location map. Follow film-shot-generation's scene-pictures reference: a wide for every scene, its opening scene picture cut from that wide, and a room map where the scene cuts between walls. Reuse one map per place and time of day; a single outdoor background uses its wide as its geography. Record the URLs in scenes[].pictures and the geography line in scenes[].geography.

Compile and verify the required pictures, quote them together with SPEND (quote it in one line, then get the yes with your host's own question tool, such as AskUserQuestion, "Yes, go ahead" / "Yes, up to N credits for this task" / "No", never as a typed question; only where the host has no such tool, those options numbered in one message; unless a spend limit the user stated covers it: N credits or an explicit no-limit ("as much as needed"), never inferred from words like "go ahead". Within it, ask nothing: quote it in one line ("About 8 credits; 12 of 20 left after this"), call it, then say what it spent and what remains, keeping a running total that includes failed calls, and stop to ask before the first call that would cross it. A money amount is asked once to restate it in credits; in unattended pace, never ask: call only with a consent or a stated spend limit already recorded, otherwise write the quote to pendingSpends and stop, as picsart-video-rules says), then generate the independent pictures in a batch. Check each before its board review, including the cast's faces, opening positions and room geometry. Use picsart_asset_approve for the location references and apply references/board-replies.md. Reuse approved pictures already in the record. A real app's screen is built as a scene layer from the user's screenshots or recording, never generated.

No still test. The scene pictures are where the cast is checked, as film-cast-design says. A cast or prop correction goes through that asset's conveyor only, under its own SPEND.

The gate into shooting: the story and shot list approved, every scene in a run has approved location references, and film-cast-design's gate into a run holds for all its tags, read back from picsart_asset_list. The opening scene picture starts the scene's first run; later runs use film-shot-generation's continuity references. No storyboard panels, voiced slideshow or animatic approval is required in the default flow.

Only when the user requests a storyboard or animatic

Record the explicit request in requestedPreview (kind: "storyboard" or "animatic", status: "requested") and load film-animatic for that work. A request for a storyboard uses its panel method only: quote and make the requested panels, check them, and review them on picsart_asset_approve. It does not buy voices, music or a timed animatic. A request for an animatic uses its timed-preview method and voice-over-generation for its sound; quote the needed pictures and sound with SPEND (quote it in one line, then get the yes with your host's own question tool, such as AskUserQuestion, "Yes, go ahead" / "Yes, up to N credits for this task" / "No", never as a typed question; only where the host has no such tool, those options numbered in one message; unless a spend limit the user stated covers it: N credits or an explicit no-limit ("as much as needed"), never inferred from words like "go ahead". Within it, ask nothing: quote it in one line ("About 8 credits; 12 of 20 left after this"), call it, then say what it spent and what remains, keeping a running total that includes failed calls, and stop to ask before the first call that would cross it. A money amount is asked once to restate it in credits; in unattended pace, never ask: call only with a consent or a stated spend limit already recorded, otherwise write the quote to pendingSpends and stop, as picsart-video-rules says), then validate, check and REVIEW the preview as that skill says. Reuse the approved cast and location references in either case.

The request chooses the work, not permission for an unquoted spend. Apply SPEND (quote it in one line, then get the yes with your host's own question tool, such as AskUserQuestion, "Yes, go ahead" / "Yes, up to N credits for this task" / "No", never as a typed question; only where the host has no such tool, those options numbered in one message; unless a spend limit the user stated covers it: N credits or an explicit no-limit ("as much as needed"), never inferred from words like "go ahead". Within it, ask nothing: quote it in one line ("About 8 credits; 12 of 20 left after this"), call it, then say what it spent and what remains, keeping a running total that includes failed calls, and stop to ask before the first call that would cross it. A money amount is asked once to restate it in credits; in unattended pace, never ask: call only with a consent or a stated spend limit already recorded, otherwise write the quote to pendingSpends and stop, as picsart-video-rules says) to every new charged operation. On approval, set requestedPreview.status: "approved" and continue to shooting; no second shot-selection question is added. Storyboards and animatics are never required, recommended as another gate, or created on a resume without that explicit request.

6. Shooting

Shoot the approved shot list directly, respecting any explicit exclusions the user already made. Do not ask them to choose the shots again. Use the approved scene-opening images and continuity references as film-shot-generation says; never require one generated panel per shot.

Recheck capacity before the forecast or quote, including on resume. Reconcile the approved shot cards, current model limits and planned runs; show the generation count and each run's seconds against the ceiling. A batch with six items is still six generations. Do not quote mergeable short runs. If correcting them changes approved run boundaries, show that correction in the existing approval/quote step; keep the user's explicit splits, shot choices and seconds. Recompile the corrected runs and their joins, and replace the old quote. Never repack a submitted or completed job into another paid generation; keep its handle and existing footage.

Before any run is quoted, a shot that film-shot-generation flags for a palette arguing with the film's scheme gets one short ASK (with your host's own question tool, such as AskUserQuestion, so the options are buttons the user taps, the recommended one first; only where the host has none, the options numbered in one message; in unattended pace, never ask: write it to the project record's openDecisions with the answer you assumed, as picsart-video-rules says) with its options.

  1. The forecast, once, before the first video: runs × each run's seconds at the model's draft rate from picsart_preflight, plus a margin for repairs, and the finish beside it. It is an estimate, not a quote.
  2. Shoot every run as a draft, in one batch, as film-shot-generation says: the model's draft setting or its lowest resolution, and no test run first. Every independent run is compiled, verified and sent together, one call per run: one quote for the batch with the finish priced beside it as an estimate, one yes, one monitor. The yes buys the drafts only. A run that waits for an earlier take (its frames are the references) is its own SPEND once that take lands and passes your check.
  3. No film so far. While takes run, never assemble, cut, open or show the film so far: a landed take waits for the others. As each take lands, find its actual cuts and judge it as film-shot-generation's takes reference says. That is your technical check, not acceptance: every take stays a candidate until the user approves the cut (7). No question between the takes and the cut.
  4. Dialogue is judged in the cut's REVIEW, by ear. Say that the words were not checked against the script by a tool.

Shooting is done when every take has landed and passed your check, with its actual cuts, prompt, quote and handle in the record. No take is accepted yet.

7. The cut

When the last take lands, film-cutting assembles the film once: every candidate take in story order with hard cuts, every generation one clip exactly as generated: no split at a run's internal cuts, no trims, no transitions, padding or freeze frames. The user's own trims and splits are kept and never recomputed. Check it with quality-review, its cold watch included (quoted first with its own SPEND (quote it in one line, then get the yes with your host's own question tool, such as AskUserQuestion, "Yes, go ahead" / "Yes, up to N credits for this task" / "No", never as a typed question; only where the host has no such tool, those options numbered in one message; unless a spend limit the user stated covers it: N credits or an explicit no-limit ("as much as needed"), never inferred from words like "go ahead". Within it, ask nothing: quote it in one line ("About 8 credits; 12 of 20 left after this"), call it, then say what it spent and what remains, keeping a running total that includes failed calls, and stop to ask before the first call that would cross it. A money amount is asked once to restate it in credits; in unattended pace, never ask: call only with a consent or a stated spend limit already recorded, otherwise write the quote to pendingSpends and stop, as picsart-video-rules says) unless every call is quoted at 0 credits; on a no, the review line says "no cold watch yet"), then REVIEW the whole cut, unasked and opened cold: this is the film's review. Once the user has answered, say what your check observed in each take. The user's approval of the cut accepts every take in it, and its shots become final.

Notes route by meaning, as film-shot-generation says: an edit, an extend or a re-roll. A note it flags as ambiguous gets one short ASK (with your host's own question tool, such as AskUserQuestion, so the options are buttons the user taps, the recommended one first; only where the host has none, the options numbered in one message; in unattended pace, never ask: write it to the project record's openDecisions with the answer you assumed, as picsart-video-rules says) between its two routes. Each is its own SPEND, shot as a draft like the takes: a yes to a run is not a yes to its repair. A note on a seam is not a licence to buy footage: film-cutting's join repair diagnoses first. A bridge between two runs is its own SPEND too, reviewed as film-cutting's join repair says. Apply what the user points at, re-cut once and open it again, until they call it final. After that, change no picture unless they ask.

8. Finishing

In this order, as film-finishing says: the finish of the draft takes, then colour, then sound.

  1. Finish the approved cut, as film-finishing says: from each draft's finalizeDraft handle where it has one, otherwise an upscale or the approved run generated again at the finish settings. It is its own SPEND (quote it in one line, then get the yes with your host's own question tool, such as AskUserQuestion, "Yes, go ahead" / "Yes, up to N credits for this task" / "No", never as a typed question; only where the host has no such tool, those options numbered in one message; unless a spend limit the user stated covers it: N credits or an explicit no-limit ("as much as needed"), never inferred from words like "go ahead". Within it, ask nothing: quote it in one line ("About 8 credits; 12 of 20 left after this"), call it, then say what it spent and what remains, keeping a running total that includes failed calls, and stop to ask before the first call that would cross it. A money amount is asked once to restate it in credits; in unattended pace, never ask: call only with a consent or a stated spend limit already recorded, otherwise write the quote to pendingSpends and stop, as picsart-video-rules says), after the cut is approved, asked as film-finishing's one question: finish the approved takes, or keep the drafts. Then the cut is re-assembled at the finish size with the user's trims.
  2. Colour: ASK (with your host's own question tool, such as AskUserQuestion, so the options are buttons the user taps, the recommended one first; only where the host has none, the options numbered in one message; in unattended pace, never ask: write it to the project record's openDecisions with the answer you assumed, as picsart-video-rules says) with three or four looks that suit the film, from the look list in picsart_scene_capabilities (sections: ["looksCompact"]). Applying one is free. When the takes look glossy or plastic, as generated footage often does, one option is the realism grade, applied as the realistic-color-grade skill says.
  3. Music: offer it once, only after the cut is final, and never bare: ASK (with your host's own question tool, such as AskUserQuestion, so the options are buttons the user taps, the recommended one first; only where the host has none, the options numbered in one message; in unattended pace, never ask: write it to the project record's openDecisions with the answer you assumed, as picsart-video-rules says) with the brief drafted as the background-music skill says, and three exits: generate it, change the brief, or skip. A changed brief is redrafted and asked again; a skip is settled for the session. A yes is a SPEND, as picsart-video-rules says. Lay the bed as film-finishing says, then REVIEW the film with it, on how it sits under the dialogue. A note on its level moves it as the background-music skill says; a different track goes back through the background-music skill's brief, with a new SPEND. A film going to festivals or dubbing gets film-finishing's sound-team handoff offered once.

9. Deliver

DELIVER the final cut once, as picsart-video-rules says. Any further master from film-finishing's delivery list is confirmed with one ASK before the first is exported. A master in another shape is reframed as the footage-reframing skill says. Then film-finishing's archive and disclosure: give the user the disclosure text (every model used, per stage) and record it, and say generated music's licence is theirs to check before a festival or a commercial use.

When something fails

  • Out of credits: zero retries. Say which step is paused and what it needs against the balance picsart_credits reads. Keep the record and every job handle. Continue only when the user says their credits are there; never poll the balance.
  • A dropped call is not a failed generation: follow its job as film-shot-generation says.
  • A board that errors on a value gets the value fixed and the board sent again. A board the host cannot draw gets the host notes' fallback, and the gate still holds.
  • Anything else a stage cannot run without gets one identical retry, then one plain line: what you tried, what blocked it, what you need. Never reword and refire.
  • An outage is status, not a decision. Keep going on everything free, keep the blocked item in openDecisions, and try again next turn. Never offer a choice the rules already answer, such as giving up a lead's consistency to dodge a temporary outage.

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.