qa-review

name	qa-review
description	Review a generated image/keyframe for brand fit, accidental in-frame text, claim/restriction violations, subject consistency, aspect, and quality artifacts BEFORE committing budget. Use after a preview/keyframe and before finalizing or generating clips. Keywords - qa, review, check, on-brand, consistency, looks right, before publishing.
allowed-tools	Bash Read Write Edit Glob Grep

Output QA Review

Catch off-brand colors, accidental text in frame, restriction violations, character drift, wrong aspect, and artifacts before you pay for the full render or the video clips. Cheapest place to fail is a 512px preview / a $0.067 keyframe — not a $14 video.

Every workflow function is exposed as a CLI command:

node workflows/cli.cjs <command> '<json-args>'   # node workflows/cli.cjs list shows all commands

---

When to run

• On preview options (generateImageOptions) before finalizeImage
• On every storyboard keyframe before generating clips
• On a product/character shot before registering it as a locked asset
• On any hero/ad image before it ships

---

How to use

Pull the brand facts from brand.md (colors, tone, restrictions.never_say) and the expected subject from the registry/plan, then review:

# brandColors from brand.md; restrictions from brand.restrictions.never_say; expectedSubject is the locked character
# checks defaults to all: ['brand-colors','unwanted-text','claims','consistency','aspect','quality']
node workflows/cli.cjs reviewOutput '{"imagePath":"projects/{name}/output-contents/storyboard/scene-01-struggle.png","brandColors":["#2563EB","#10B981"],"brandTone":"professional, warm","restrictions":["guaranteed","cure","risk-free"],"expectedAspect":"9:16","expectedSubject":"late-20s East-Asian woman, beige knit, gold hoop earrings"}'
# data: { pass, score, issues:[{check, severity, detail}], summary }

---

Acting on the report

pass === true  → proceed (finalize / generate clips / register).
pass === false → there is an ERROR-severity issue. STOP and show the user:
   • the issue list (check + detail)
   • offer to regenerate with a corrected prompt addressing each error
   • re-review the new output before spending
warnings/info  → surface them, but the user may accept and proceed.

Always show the user the issues and the score — don't silently pass or block. A failed QA on a cheap preview is the system working: it just saved a wasteful expensive run.

Reviewing a video? Extract a representative frame first (ffmpeg -y -i clip.mp4 -vf "select=eq(n\,0)" -frames:v 1 frame.png) and review that.

---

Typical loop

generateImageOptions → reviewOutput(each) → user picks a PASS → finalizeImage → reviewOutput → ship
generateStoryboard   → reviewOutput(each keyframe) → fix fails → approve → generate clips

Videos → reviewVideoOutput (2026-07-05)

Video QA is one call now — no manual frame extraction: node workflows/cli.cjs reviewVideoOutput '{"videoPath":"…","expectAudio":true,"expectedDurationSeconds":N,...}' Samples frames + checks container (duration/audio). For kinetic/typography content pass checks WITHOUT "unwanted-text"/"claims". Run after EVERY video generation; fix what frames[] flags before packaging.