Prompt validator

prompt-validator

Deterministic pre-spend gate: collect every prompt a video run is about to send (all scenes, all blocks), cross-check against each other and against the run's locked facts (character lock, vocal map, model capabilities), block with a specific error on contradiction. Skills are deterministic — a multi-scene batch fails by contradiction, not drift: two prompts that cannot both be true of the same video (AI-council convergence, 2026-06-06). One blocked batch costs a re-edit; one rendered contradiction costs the whole spend.

When to use

• Immediately before the unified storyboard/cost preview gate in /video:from-song Step 8 / /video:from-script — after every scene's eight-block prompt exists, before anything is shown for approval or sent live.
• After an operator hand-edits scene prompts (/video:scene re-runs) — re-validate the batch, not just the edited scene.

Do NOT use when:

• Only a single scene exists (nothing to contradict) — per-scene checks in video-director cover internal consistency.
• Pure preview with no approval intent — validate when the batch is about to be gated, not on every draft keystroke.

Inputs

• <project>/script.md — every ## Scene N (duration, mood, action, camera, dialogue).
• Every rendered scene prompt (eight-block image/video prompts from video-director / blueprint JSON), in scene order.
• <project>/character.json — when present, the locked identity (subject, palette, wardrobe, prop, seed).
• <project>/vocal-map.json — when present, the signed-off line → singer assignments.
• The chosen model's capability entry (capability --model <id>).

Procedure

1. Collect the full prompt set

Read every scene's prompt blocks into one table: scene → {style, subject, environment, action, camera, lens, lighting, mood, duration, aspect, dialogue/singer}. Prompt missing or unparseable → block: scene N: prompt missing/unreadable — an incomplete batch already contradicts the script.

2. Style consistency

style block must agree across all scenes — one look per video unless the script names an intentional break (e.g. "dream sequence").

• Conflicting render families (photorealistic vs cel-shaded / claymation / watercolor) with no scripted break → style mismatch: scene 3 'photorealistic' vs scene 7 'cel-shaded'.
• Conflicting era/grade tokens (1970s film grain vs clean digital HDR) → same error shape, name both scenes + both tokens.

3. Character consistency (character mode)

With character.json present, every scene naming the subject must match the lock:

• Wardrobe / palette / prop drift (red coat locked, scene says blue jacket) → character mismatch: scene 5 'blue jacket' vs lock 'red coat'.
• A scene inventing a second recurring subject the script never cast → flag — uncast identities break continuity AND the likeness gate.
• Lip-sync ownership: a scene with a dialogue: line must name the vocal map's singer for that line — wrong mouth on wrong words is the render-money-burning failure (media-sync-ground-truth). A "?" singer with a lip-sync scene → block until resolved.

4. Physics + continuity

• Day/night or weather flips inside one contiguous song section with no scripted transition → continuity: scene 4 'noon sun' → scene 5 'starry night' within the same chorus.
• Impossible camera/action pairs (locked-off tripod + whip pan), (underwater + desert dust) → name scene + both tokens.
• Duration / aspect vs. model envelope: any scene outside [min_duration, max_duration] or with an unlisted aspect → block with the violated bound (defense-in-depth after song-to-script Step 1 — hand-edited prompts re-open this hole).

5. Verdict

• Zero findings → emit prompt-validator: N scenes, no contradictions, hand the batch to the preview gate.
• Any finding → block the gate. Report every finding (not just the first), each as class: scene(s) + conflicting values + fix hint. Operator (or calling command) fixes prompts, re-runs the validator — never "approve anyway" past a contradiction; an intentional break gets scripted (named in script.md), which legitimizes it on the next pass.

Output format

1. One line per finding — class: scene(s) + conflicting values + fix hint — and report every finding, never just the first.
2. Terminal verdict: line: OK with the scene count, or BLOCKED (N contradictions) — fix and re-run; a blocked verdict never hands the batch to the preview gate.

prompt-validator: 12 scenes checked
✅ style: one look (neo-noir, consistent)
❌ character: scene 5 'blue jacket' vs character.json 'red coat' — align scene 5 or update the lock
❌ lip-sync: scene 9 dialogue assigned to 'Freya' but vocal-map line 24.5-27.8 is singer '?' — resolve the singer first
verdict: BLOCKED (2 contradictions) — fix and re-run

Gotchas

• An intentional style break is scripted, not waved through. Only a break script.md names is accepted; "operator said it's fine in chat" does not survive to the next run.
• Re-validate after every hand-edit. One edited scene can contradict five untouched ones; the batch is the unit, not the scene.
• verified: false capability entries still gate. Blocking on a documented-best-effort bound beats rendering past a real one; the report names the unverified flag — override by editing the manifest, never by skipping the check.

Do NOT

• Do NOT auto-fix prompts — report and block; the fix is the operator's (or calling skill's) edit, visible in the diff.
• Do NOT validate only changed scenes — always the full batch.
• Do NOT soften a contradiction into a warning — contradictions block; only consistency passes.
• Do NOT duplicate video-director's per-scene craft checks — this skill owns cross-scene contradiction only.

See also

• /video:from-song — invokes this before the unified storyboard/cost preview gate (Step 8).
• video-director — produces the prompts this skill cross-checks.
• character-consistency — the lock this skill enforces against.
• song-to-script — upstream timing + vocal-map source.