App Farm — concept-anchored creative factory for paid-ads platforms

Build a farm of small Cursor-coded apps, one app per creative concept type (close-up emotional female testimonials, packshot grids, talking-head explainer, etc.). Each app reads a row from an Airtable variable matrix (persona × hook × geo × language × ad set), generates keyframe images via nano-banana (Gemini 3 Image Pro), strings them into video via Kling AI or Veo, assembles with packshot / voice / music, and ships via the Meta Marketing API straight to ad sets. A daily script pulls Hook Rate + Hold Rate per asset, flags underperformers, and on operator approval kills them. Concept apps stay small and disposable; the matrix and the kill-script are the durable assets.

Reported outcomes from operators running this loop on a consumer subscription app at ~10M users: "record month for purchases and revenue within a week" after switching from copy-paste-plus-variation to per-concept apps. One agency partner producing creative for clients reports 2,000 complex creatives for one client in one weekend; another client's request for "just 8 concepts" decomposed to 1,872 distinct assets (60 images + 60 video assets + 15 talking heads + 15 voices + 15 slide shows + 4 utility grids + 1 packshot + 15 combined videos per concept, plus ~20% garbage). Without the per-concept-app structure, none of that volume reaches the auction.

1. Stop making cosmetic variations. Meta and similar platforms deduplicate server-side via perceptual hashing. Hook swap, button color change, even a different opening line on the same shot — these collapse into one concept in the ranking and one of them wins. Variations must be different concepts, not cosmetic edits.
2. Define your concept catalog. Audit which creative concept types actually work for your category (close-up testimonial, packshot grid, talking-head explainer, animated explainer, lifestyle b-roll). For each, you'll build one Cursor app.
3. Airtable as the variable matrix. One table per concept catalog. Columns: persona, voice, location, hook, language, ad set, status. Each row is a generation job. Marketing operators add rows; the apps consume them. Airtable's the user-visible spine — engineers don't need to touch it for ops.
4. One Cursor-coded app per concept. Each app: read N rows from Airtable, generate keyframes via nano-banana, run keyframes through Kling AI or Veo to get clips, layer packshot + voice + music, write the result back to Airtable with status and a thumbnail. Apps are small, single-purpose, disposable. Don't try to build a generic super-app — every operator who tried that ended up with a worse Airtable.
5. Upload via Meta Marketing API directly. Skip Ads Manager UI. Script the upload — batched into 50 creatives/day across iOS/Android campaigns is the working volume one operator runs. One missing API toggle ("App store details (Off)") still needs manual handling; everything else is scripted.
6. First-frame rule for talking heads. If the starting frame has a closed mouth, nano-banana / Imagen 4 will render unnaturally white "toilet-tile" teeth on the first open. Force the starting frame to include visible teeth in the prompt.
7. Daily kill-script. Pull spend + Hook Rate + Hold Rate per asset 24h after launch. Flag underperformers. After operator approval, the script switches them off. Hook Rate + Hold Rate dominate the ranking — bad hook in the first 2 seconds, nothing downstream saves it.
8. Run an internal self-check loop on the generators. Each concept app should pass its own output through a critic agent (matches brand? matches hook brief? face composition acceptable?) before queueing for upload. By the 3rd iteration of a concept app, the operator reports it produces shippable creatives without human review.

Verbatim concept-app architecture:

Per concept app:
 read N rows from Airtable (persona × hook × geo × language × ad set)
 generate keyframes: nano-banana
 string into video: Kling AI or Veo
 assemble: packshot + voice + music
 upload: Meta Marketing API → ad set
 write back: Airtable row → status + asset URL

Daily kill-script outline:

# cron: 0 9 * * * (daily 9am, 24h after the previous day's batch launch)
for asset in fetch_active_assets():
 stats = meta_api.get_asset_stats(asset.id, last_24h=True)
 if stats.hook_rate < HOOK_THRESHOLD or stats.hold_rate < HOLD_THRESHOLD:
 flag_for_kill(asset, reason=stats)
notify_operator(flagged_list)
# operator clicks approve; script switches off after confirmation

• Cosmetic variations get deduped. Hook swap or button color is not a new concept to the ranker. Variations must be concept-level different.
• Toilet-tile teeth on closed-mouth starting frames. Force visible teeth in the first-frame prompt.
• AI disclosure labels drop CTR ~33%. Research cited in source. When platforms auto-label AI-generated content, pre-plan budget reallocation. Avoid AI-impersonating-real-people testimonials entirely — FTC-prohibited and hit with $1,000–$5,000/violation state fines.
• ComfyUI is overkill for ad creative. Stay on nano-banana + Kling/Veo. Reserve ComfyUI + Unreal Engine 5 + Wan2 + VACE for film-grade VFX or brand-conformant pipelines where you can't risk text-to-video drift.
• Don't build a single super-app. Every operator who tried collapsed back to per-concept apps. The matrix lives in Airtable; the apps stay disposable.
• TikTok-virality doesn't predict Meta. Different ranking weights; use organic TikTok as directional signal at best, not as the A/B filter.