promo-critic
Skill1 starsChanged 45 days agoMITView on GitHub
Critique a rendered promo or explainer video from the product promo harness like a harsh motion director (`promo critique` evidence: contact sheets, seam strips, phone sheet, sync and loudness measurements), hold it to the user's hard quality rules, score it, and drive the fix-and-rerender loop up to `promo ship`. Use when reviewing, critiquing, scoring or QA-ing a rendered promo video, or when a promo-video run reaches its critic gate.
In this repository4 other files it gives its agents.
- promo-captionsSkill
- promo-scriptSkill
- promo-videoSkill
- promo-voiceSkill
The file
We split this skill into the rules it gives an agent
6 rules. About 0 are common, 3 are distinctive, and 3 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈3
Distinctive · estimate
This author's own choices. Worth a look.
3
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Very specific: line 22
Read review.md. Fix the three worst problems, re-render only the affected shots (promo scenes --shot <id>), promo assemble, and start the …
- Very specific: line 32
No blur, anywhere: No blur filter, depth of field or motion blur on anything, no blurred background, no blur-in entrance, no …
- Very specific: line 40
App truth: Never invent UI, copy, numbers, citations or identifiers.
- Distinctive: line 45
Fonts are vendored: No CDN or network at render.
- Distinctive: line 47
No CSS outlines or focus rings.
- Distinctive: line 51
Every score is 8 or more, every pack checklist item passes, every automatic check passes and no hard rule is broken.
≈ “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
promo-critic
One round
promo verifyfirst (determinism: the same frame whatever the seek order). Thenpromo critique(orpromo critique --video renders/draft/animatic.mp4for a structure round on the animatic, whosetechcheck fails on the half-size, 30 fps draft by design). It writescritic/<n>/evidence/(contact sheets on the beats or every 0.5 s, a strip at every seam and at the fastest action, a phone-size sheet, a master sheet, the spectrogram,loudness.json,sync.json,tech.json),critic/<n>/summary.jsonandcritic/<n>/PROMPT.md. It exits 1 when an automatic check fails.- Fix failed automatic checks first (
tech,loudness,sync_sfx,sync_vo,vo_wer,claims,determinism,hf_lint); they are facts, not taste. Re-render, re-assemble, critique again. - Dispatch a fresh subagent that has not seen the plan or the code, on the most capable model. Give it only:
the path to
critic/<n>/, the instruction to readPROMPT.md, open every image inevidence/and check every frame against the hard rules below (paste them verbatim), and to writereview.mdandscores.jsonnext toPROMPT.md. Do not summarise the video for it. - Read
review.md. Fix the three worst problems, re-render only the affected shots (promo scenes --shot <id>),promo assemble, and start the next round.
Hard rules (any frame that breaks one fails the round)
- Readability. Key text is 22 px or more on screen at 1080p on its beat. The camera pushes in on the element
the line is about and frames whole components; no word is cropped. The evidence sheets are scaled down (22 px is
about 4 px on the 360 px-wide phone sheet), so before dispatching, add full-size frames of the doubtful beats to
evidence/(ffmpeg -ss <t> -i renders/master.mp4 -frames:v 1 critic/<n>/evidence/full-<t>.png), or measure with a readability script that reads every visible text node's rendered height. - No blur, anywhere. No blur filter, depth of field or motion blur on anything, no blurred background, no
blur-in entrance, no
backdrop-filter. Fail any blur you see. Competitor names are kept out of frame by the beat or the crop, never blurred. - No white inner shadow, glow, feathered edge mask, vignette or grey scrim on any frame.
- Cursor. A pointer (hand) cursor over clickables. Each click lands on the real target (button, row or field, never a label or panel title). The cursor arrives 0.3-0.5 s before the click.
- Zooms. Smooth and accurate, landing framed on the exact element; nothing important clipped (no half dialogs, no cut-off documents or panel names); the app window is big enough to hold the content whole.
- App truth. Never invent UI, copy, numbers, citations or identifiers. Values on screen are computed from the
app; a lookup shows only real metadata. Compare against
assets/capture/and the app itself. - No dead frames. Something new every 2-4 s; never a hold on a sparse or empty grid.
- Mascot or characters, when used. Hero scale when they are the shot, natural motion, no anatomy glitches (extra limbs, lost line work, foot slide).
- Fonts are vendored. No CDN or network at render; a substituted font is a failure.
- Public copy went through the
humanizerskill: no em dash, no "not X but Y", no hype words. - No CSS outlines or focus rings.
Pass bar
Every score is 8 or more, every pack checklist item passes, every automatic check passes and no hard rule is broken. At least 2 rounds, at most 5. After round 5, stop and report what still fails; never claim success you did not measure.
The ship gate
promo ship accepts only the latest round, and only when its summary.json has ok: true, its scores.json
has pass: true, and the video it judged is this master: the round records the SHA-256 of what it judged, so an
animatic round or a re-assembled master does not count. The last round is always on the final master.
promo ship --force ships anyway and writes that it did into the run README; use it only when the user says so.
A film silent by design passes loudness and sync_sfx as "n/a": its shot list plans no VO lines and no SFX
placements, and its pack's music mode is none. Anything planned but not made fails and names the step to run.
Honesty
Audio is checked by measurement (LUFS, true peak, sounds found at their planned events, ASR on the mix), not by ear: say so. A critic that praises everything is not working; if round 1 scores all 9s, ask the subagent to look again at the seam strips and the phone sheet.
Companion skills for the fixes: hyperframes-agent-render-gotchas (frames that differ by seek order, fonts,
frozen renders) and ffmpeg-filtergraph-mix-master-gotchas (loudness, true peak, ducking, the master chain).
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.

