Generating AI Video Ads

Minimum input needed: product, audience, platform+length, format (or let the skill recommend one).

Product: Reusable silicone food wraps
Audience: Eco-conscious home cooks, 25-45
Platform/Length: Instagram Reels, 30s
Format: (not specified → skill recommends)

Skill recommends: B-Roll-Driven (primary) or Realistic UGC (alternate) — physical product, demo-friendly, low lip-sync risk. Proceeds with B-Roll-Driven.

Output (in order): assumptions → script w/ hook variants → timestamped shot table → style bible → image prompts → motion prompts → format extras (if any) → toolchain → assembly checklist → QA report → editing notes.

Progress:
- [ ] 1. Intake & brief lock (one clarifying round max)
- [ ] 2. Format selection / recommendation + fit check
- [ ] 3. Script (hooks + body, framework-driven)
- [ ] 4. Shot list + timestamps (sum must match length exactly)
- [ ] 5. Style bible + visual board (locked, verbatim-reusable)
- [ ] 6. Image prompts (one per shot, target-tool syntax)
- [ ] 7. Motion prompts (one per shot, single action + camera move)
- [ ] 8. Format extras (lip-sync sheet / lyrics+beat map / premise+twist)
- [ ] 9. Toolchain (primary + fallback per stage, budget-aware)
- [ ] 10. QA pass (timing, claims, IP, AI-failure points, disclosure)
- [ ] 11. Deliver: assembly checklist + editing notes + regenerate-first list

1. Intake & Brief Lock

Required: product, audience, platform/length, format (or auto-recommend). Everything else optional → convert missing optional fields into one-line stated assumptions. Ask at most one clarifying round, and only for missing required fields.

If format is missing: score candidates against product type + audience + budget, state top pick + one alternate, proceed with top pick.

2. Format Fit Check

Cross-reference format against product type (see products/product-type-rules.md). If it's a poor fit (e.g., Claymation for a luxury watch, Crochet for B2B SaaS), warn in one line and suggest an alternative — but follow the user's explicit choice.

Route to the matching format module (formats/*.md); fall back to the shared core pipeline for formats without a dedicated file.

3. Script

Pick a framework by format/product (see table below), write 2-3 hook variants (question, bold claim, pattern interrupt, POV, curiosity gap, confession, visual shock, social proof — pick 2-3 varied types), then the full voiceover/dialogue at ~2.5 words/second.

4. Shot List + Timestamps

One action per shot, 2-5s each, total time = exact platform length. Columns: shot #, timestamp, voiceover/dialogue, shot type, on-screen text. Song ads: one shot per lyric line, product hero shot on the chorus.

5. Style Bible + Visual Board

Lock: character/presenter/object design, palette, lighting, environment, camera lens feel. This text block gets pasted verbatim into every image/motion prompt for the ad — never paraphrased per-shot.

Critical rule: never name a studio, creator, franchise, or celebrity. Translate instead:

• "Pixar-style" → "soft 3D animated feature-film look, subsurface skin, rounded shapes"
• "Zak D Film style" → "cinematic VFX-heavy dramatic short-film look, volumetric haze, epic scale, photoreal textures"
• Any named creator/style → descriptive visual language only

6. Image Prompts

Formula: subject + action + setting + style + camera + lighting + mood + aspect ratio + negatives. Write in the target tool's syntax (Midjourney, Flux, Ideogram, Nano Banana, etc.) — if user didn't name a tool, default to the format's recommended primary and note the fallback syntax too.

7. Motion Prompts

One motion + one camera move per clip; state what stays still. Prefer image-to-video from a locked keyframe over text-to-video for anything that must stay on-model. Respect the chosen tool's clip-length limit.

8. Format Extras

• Talking Head / Talking Object / Song Ad: lip-sync sheet (audio file per line, mouth-open moments, clip order) — generate audio first, drive lip-sync from it, one speaker per clip.
• Song Ad: original lyrics (never imitate existing songs), genre, BPM, chorus hook, lyric-to-shot map.
• Cinematic VFX: one-line premise + twist reveal, VFX difficulty rating per shot.
• All formats: regenerate-first list of riskiest shots.

9. Toolchain

Pick by job, not brand, unless user names a tool (then use that tool's syntax). One primary + one fallback per stage (image, video, lip-sync, voiceover, music, editing, upscale). Each tool entry carries a "last verified" date; flag anything older than ~60 days and say so rather than guessing.

10. QA Pass

• Timestamps sum to exact target length
• Word count fits ~2.5 words/sec pacing
• Product visible in ≥50% of shots
• Claim checker: no health/income/"#1"/guarantee language unless supplied (extra caution: supplements, finance, alcohol, kids)
• No studio/creator/franchise/celebrity names in any prompt
• AI-disclosure note added if platform/content type requires it
• Format-specific AI-failure-point check (see table below)

11. Deliverables (fixed order)

1. Assumptions (1-2 lines)
2. Script with hook variants
3. Timestamped shot table
4. Style bible + visual board
5. Image prompts (copy-paste, one per shot)
6. Motion prompts (copy-paste, one per shot)
7. Format extras (if applicable)
8. Toolchain (primary + fallback per stage)
9. Assembly checklist (generate → retime → caption → export)
10. QA report + regenerate-first list + credit-saving plan
11. Editing notes (cuts, overlays, music/SFX, glitch-hiding)

Example 1 — Format given Input: Product: protein cookie brand "OatBoss"; Audience: gym-goers 20-35; Platform: TikTok 15s; Format: Realistic UGC Output: Assumptions note (tone=casual, CTA="link in bio"), hook-body-CTA script with 3 hook variants, 5-shot table (3s each) alternating selfie/demo/result, style bible (creator look, room, phone-grain realism), 5 image prompts with natural-light/phone-camera-grain keys, 5 motion prompts with light handheld shake, toolchain (primary: GPT-image + Kling native audio; fallback: Midjourney + Runway), QA report flagging shot 4 (hands holding product) as regenerate-first.

Example 2 — No format given Input: Product: kids' building-block app; Audience: parents of 5-9 yr olds; Platform: YouTube Shorts 30s Output: Skill recommends 3D Animated (primary) over Talking Object (alternate) — kids/toys category, child-likeness risk avoided via animated format. Proceeds with 3D Animated: mini-story script (setup/conflict/payoff), character sheet locked first, style bible uses "soft 3D animated feature-film look" (no studio name), 6 shots, prompts with subsurface-skin/rounded-shapes keys, gentle character motion, toolchain for stylized-character consistency tools.

Example 3 — Style name requested Input: User says "make it look like a Pixar movie, Zak D Film style for the VFX shot" Output: Skill silently translates — style bible reads "soft 3D animated feature-film look, warm volumetric lighting, rounded character proportions" and, for the VFX shot if cross-format, "cinematic VFX-heavy dramatic look, volumetric haze, epic scale, photoreal textures." No prompt contains "Pixar" or "Zak D."