agentleFS
PicsArtPicsArt/gen-ai-skills/skills/quality-review/SKILL.md

quality-review

Skill7 starsChanged 20 days agoMITView on GitHub

Check work before the user sees it - a scene validated, linted and looked at in contact-sheet frames, then fixed and checked again; a draft reviewed by critics, after a cold watch against its viewer track when one runs; the brief graded; an export checked against its scene. Use after a script is written and before it is built or paid for, after building or changing a scene, before REVIEW or an export, before handing over an export, and at each decision a use case names.

8 other projects use a version of this.

Adopted by

Most recently taken up by

The file

We split this skill into the rules it gives an agent

8 rules. 0 are common, 6 are distinctive, and 2 are very specific to this project.

  • 0

    Common

    Most of the 6 edited versions we read keep these.

  • 6

    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
  1. Very specific:

    when the use case built it in a style, run the visual-style skill's adherence.py on it, and check the frames against that skill's list of …

    line 29
  2. Distinctive:

    every second has something new: no stretch over 3 s where nothing moves, appears or changes, unless it is a hold the use case asked for.

    line 34
  3. Distinctive:

    not one move everywhere: when every beat has the same push or the same entrance, vary what the use case's skills allow.

    line 44
  4. Distinctive:

    The brief, item by item: When the use case has a brief (the message, the destinations, the length, the hook, the call to action, the …

    line 53
  5. Distinctive:

    The cold watch: Before the critics, and before the user sees a draft that has a viewer track, three model viewers who know nothing of it …

    line 65
  6. Distinctive:

    After the cold watch, before the user sees a draft, and six simulated critics with taste each look at it through one lens: the …

    line 77
  7. Distinctive:

    Keep the scene as it is. Tell the user which feature fails, with the error and the smallest scene that shows it, and offer a way round it …

    line 123
  8. Very specific:

    Before a short is built, voiced or generated, review its script: is it worth making, what is its why, and which lines to rewrite.

    line 165

“Common” is measured: kept by at least 6 in 10 of the 6 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.

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

Evaluate

  1. Validate. A successful picsart_scene_patch has already validated its result, so call picsart_scene_validate only after changes made any other way. Zero errors; fix the warnings too.
  2. Lint the layout. picsart_scene_lint_layout over the timeline; on a long scene, narrow t0/t1 to the part in question. Each of its findings is a sustained overlap between layers that should not touch. For a pair placed together on purpose, pass allow: ["idA:idB"]. A text box is the font's line box, so a clean lint is not proof and it can flag empty space. If the tool is not in your list, or fails because the query build is unavailable, skip to the frames.
  3. Measure. For a flagged pair, or a use-case question about where something sits (inside a panel, inside a mask), picsart_scene_query_layout at that moment gives each layer's box. It reads layer geometry, not painted pixels.
  4. picsart_scene_contact_sheet at each moment the use case names. Look at every frame and answer the use case's questions about it.
  5. picsart_scene_contact_sheet once across the whole timeline — order, seams, nothing blank. Then, for every cut, its times at the cut and one frame either side (cut − 1/fps, cut, cut + 1/fps): a one-frame black gap falls between evenly spaced samples.
  6. Polish, once the frames are right. A correct video can still look machine-made, and the user notices that first:

    • when the use case built it in a style, run the visual-style skill's adherence.py on it, and check the frames against that skill's list of defaults that look machine-made;
    • the first frame is a real moment: the payoff, or the problem it solves, as the retention-hooks skill's hook rules say. Not a login screen, an empty app or a title on a plain canvas;
    • every second has something new: no stretch over 3 s where nothing moves, appears or changes, unless it is a hold the use case asked for. First run scripts/timing.py on { scene }, from this skill's folder: from the JSON alone it flags every stretch over 8 s with no change (dead_air), every block of words held under 0.3 s a word or 1 s (short_hold), and every block over 10 words, or a step label over 6 (too_many_words). Typed holds add only the typing remaining in each visible span, including after retiming or trimming. Text and caption masks in mask.clip or legacy mask.layer need rendered review; the script rejects them. Pass the measured reveals inside a clip as changes; the JSON cannot see them. Then look at the frames for the 3 s: motion inside a clip is only on them;
    • not one move everywhere: when every beat has the same push or the same entrance, vary what the use case's skills allow;
    • words inside the safe zone of the destination the use case names: run the screen-recording-layouts skill's safe_zone.py with its profile, and fix each finding. Readable at the frame's width.

    Fix each finding, or keep it and say why in the review line ("the app itself is navy, so the canvas follows it").

  7. Fix the layers and check again. Building is free.

The brief, item by item. When the use case has a brief (the message, the destinations, the length, the hook, the call to action, the steps), grade the build against it one item at a time, never as one judgement of the whole. From this skill's folder, run python3 scripts/checklist.py <file> on { brief, scene }, with the destinations as the screen-recording-layouts skill's safe-zone profiles. It turns the brief into yes-or-no questions and grades the ones the scene answers: the length, words by 0.5 s, the hook's words, the call to action at the end, the safe zone, and timing.py. Checks the scene cannot answer for a text mask become look items, alongside the remaining checks. Answer each from the frames of steps 4-6, with the frame's time or the tool's line as evidence, and run it again with your answers; an answer without evidence stays open. Fix only the items that say no, then grade again: never rebuild what passed. Its summary goes with the review.

The cold watch. Before the critics, and before the user sees a draft that has a viewer track, three model viewers who know nothing of it watch it forward, a few seconds at a time, from its frames and its heard words, each keeping only their own notes, and say what they make of it. A comparer that never saw the plan marks their words against the track. scripts/cold_watch.py writes every call, checks the marks and says whether the floor is met: every must-know hit in time by two of the three, and no wrong ending. It is the one check of whether a stranger follows the video: nobody who knows the plan, you included, can judge that. Its viewer calls cost credits unless their quote shows 0 for every one: quote them as references/cold-watch.md says and 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) before the first round; a credit or storage error stops it, never retried. It is not a measure of retention, and it says so.

The critic panel. The panel meets After the cold watch, before the user sees a draft, and six simulated critics with taste each look at it through one lens: the first-principles critic (the core problem the video solves), the radical (the concept), the interface designer (the UI on screen), the story editor (the message and the script), the growth strategist (the stop, the drops, the ending) and the art director (the look). Each sees only frames every 0.5 s, the source stills (the real UI or footage the video depicts), the words, the goal, the destination and the cold watch's gap list when one ran, and returns a verdict, problems with their times, and fixes. scripts/critics.py picks the frames and lays the notes side by side. Where critics agree it is a must-fix, weighed against the first-principles critic's core sentence; a new concept goes to the user as a choice, never applied silently. So does a shot a critic would swap in an edit of the user's footage: it is a proposal, shown to the user as a rendered A/B, never a silent change to the plan. A draft that looks different from its source is a must-fix.

The challenger. A seventh critic who does not wait for the draft. At each decision the use case names (the idea, the plan, the key frames and each image picked, the UI, and every SPEND before its quote) it asks "why this?" and tries to show that the direction is wrong before more work is spent on it. It returns keep, rethink or stop, with the reason and the cheaper alternative. It also gets the source stills, so a design that drifts from the real app is caught as it is made. Its rethink of a shot the user saw and kept is a proposal too: an A/B, never a silent swap.

The QC inspector. An edit of the user's own footage is checked from its render by an inspector who did not make it and never sees why it was made that way: camera steadiness on every used range, lips against the voice, each cut against its word, shot lengths, legible text, captions, repeated or clipped lines, the locked meaning map, and every keeper of a version the user saw: one dropped with no line the user will read fails the draft, and, when a reference's format is in the record, the reference match (scripts/reference_match.py reads it from the scene). It reports only critical and major findings, each with its method and confidence, and marks what it could not see or hear as open, never as a pass. scripts/inspection.py reads its report back.

Read thisBefore you
references/qc-inspector.mdcheck a draft of an edit of the user's footage, or a fix to one
references/cold-watch.mdwrite a viewer track
, or run the cold watch on a draft
references/cold-watch-comparer.mdmark a cold watch's accounts against the track (the comparer's prompt)
references/critic-panel.mdrun the critic panel on a draft
references/challenger.mdgo on from a decision the use case names, or show a quote

When a frame or an export fails to render. The scene validated, so the fault is in a feature the renderer refuses, and the error names it rarely. Find it before changing anything:

  1. Copy the error text exactly.
  2. Cut a copy of the scene down to the one layer at the failing time, and sample that. Then take away one feature at a time (its mask, its effects, its animation) until it renders. Each sample is one frame.
  3. Keep the scene as it is. Tell the user which feature fails, with the error and the smallest scene that shows it, and offer a way round it that keeps the look where one exists (for a mask, the screen-recording-layouts skill's nested panel). Never reshape the whole scene to guess past it.

Before a film's cut is reviewed. From this skill's folder, run python3 scripts/cut_check.py <file> on { scene, allowed?, shots?, places? }: the cut's scene, its allowed spans (the screen clips placed in it, as film-cutting says), and from the project record the shot each clip shows (keyed by clip id, with its size and angle when the shot list has them) and the place each shows. From the scene alone it finds the joins, moves and overlays a viewer reads as a glitch: two takes of one shot, or two shots of one size and angle, cut together (same-scene-cut), or a cut hidden under a quick move (hidden-take-change); different shots of one place are coverage and pass, a short dissolve between takes (ghost), a move of the picture, or of a layer covering 40% of the frame or more, that starts or stops in one frame (pop-move), an overlay gone for a frame or two (blink), a zoom over a still picture (zoom-idle), and a screen held still inside its span (screen-still). Nothing inside a span is flagged except screen-still. It prints { ok, errors, warnings, unchecked, skipped }: fix each error as film-cutting says and run it again, say each warning in the review, and check what it lists as unchecked on the frames of step 5, never calling it checked. On an animatic of held stills, add --slides: its stills hold on purpose, so the motion rules move to skipped and run on the cut, and the joins, overlays and screens are still checked.

Before you hand over an export. An export can differ from its scene: another size, a dropped frame rate, a missing sound track, a silent last second. Probe the file with picsart_media_probe, which is free, and from this skill's folder run python3 scripts/qc.py <file> on { scene, probe, audio? }: the scene you exported, the probe as it returned, and in audio only what a tool measured (loudness, true peak, the sound's length, an energy curve). It prints { ok, errors, warnings, unchecked }:

  • an error (the size, the fps or the length is not the scene's, the sound is missing, drifts or peaks above -0.5 dBTP): make its fix and export again, once. Still there: deliver, and name it;
  • a warning (loudness away from -14 LUFS, a silent tail over 1.2 s): say it in the delivery line;
  • unchecked: what no tool measured. Never call it checked.

Frames show framing, not motion: the user judges timing and easing in REVIEW.

How many rounds you get, and what to do when they run out, is in picsart-video-rules (Evaluation). Never export to "show" a draft that the user can review live; picsart-video-rules (Preview and export) says when to export.

The script, before the build

Before a short is built, voiced or generated, review its script: is it worth making, what is its why, and which lines to rewrite. A video whose footage never shows its result is dropped before it costs anything. The review runs scripts/script_review.py; the reference below has its steps.

Read thisBefore you
references/script-review.mdreview a script, or report its verdict
references/readiness-verdict.mddeclare a generated video's decisive evidence before paying for it, or say whether a candidate is ready

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.