Generate 10+ strategic ad creative iterations from a winning Meta/Facebook ad through a 4-Filter Winner Validation Gate and a 5-vector iteration system.
Give Claude a winning Meta ad and it runs the full chain — detects what's connected (Meta Ads MCP, /watch skill, Competitor Ad Spy), pulls your account data filtered to sales-objective campaigns only, derives benchmarks from your own ads, validates the candidate winner through the 4-Filter Test (Spend / Signal / Stability / Scale), diagnoses the winning DNA across 7 components, then produces 10+ iterations ordered by the Andromeda Template Rule (templates first, copy second). Refuses to iterate on fake winners.
Claude Desktop (Cowork): download creative-iteration-engine.skill → Settings → Skills → drop it in.
Claude Code:
git clone https://github.com/the-baweja/creative-iteration-engine.git ~/.claude/skills/creative-iteration-engine
Manual: clone this repo into any skills directory your Claude setup reads from.
You give it a winning Meta ad — or no winner yet, just a brand. It runs the full pipeline:
- Preflight Check + Setup Runbook — detects what's connected (Meta Ads MCP, /watch skill, Competitor Ad Spy), gives you exact install instructions for anything missing, offers fallback paths so you're never blocked
- Mode selection — Mode A (you have a winner with performance data) or Mode B (no winner, enter the category strategically via competitive analysis)
- Data import (3-tier ladder) — Meta Ads MCP → CSV upload from Ads Manager → manual paste. Always reads the highest tier available.
- Benchmark derivation — pulls last 30 days at ad level, filters to sales-objective campaigns only (no skewed averages from awareness/traffic), computes account-specific CTR / CPM / CPC / CPA / hook rate / hold rate using top-quartile median as the bar
- 4-Filter Winner Validation Gate — Spend (4-5× / 10-15× / 21-25× CPA thresholds), Signal (CPM + CTR + CPC read together), Stability (post-budget-bump hold), Scale (CPA drift). Refuses to iterate on fake winners.
- DNA diagnosis — breaks the winner into 7 components (hook, angle, proof, offer/CTA, format, tone, audience signal), identifies the 2-3 elements driving performance
- Video import — chains into the
/watchskill when the ad is a video. Downloads, extracts frames, transcribes — iterations are grounded in what's actually on screen, not guessed from the title. - Iteration generation — produces 10+ iterations across 5 vectors, ordered per the Andromeda Template Rule:
- Template/Format swap (safest) — different visual structure, same message
- Hook swap — same template, different opener
- Proof swap — different testimonial / stat / before-after / demo
- Tone swap — UGC ↔ polished, positive ↔ fear ↔ contrarian
- Angle extension (experimental) — new persuasion approach grounded in the same audience insight
- 100-Variation Mode (exceptional winners only) — if the ad clears all 4 filters, produces 100 variations across templates / hooks / proofs / tones / formats / placements / angles / contrarian tests
- Scoring + 3-wave testing roadmap — each iteration scored on a 6-factor rubric. Wave 1 validates, Wave 2 explores, Wave 3 discovers. Per-wave kill / iterate / scale thresholds anchored to your own benchmarks.
Plus a decision framework — what every test outcome teaches you and what to test next. Every result becomes strategic intelligence, not pass/fail.
Mode A — Own Winner. You have a Meta ad that's working and want 10+ strategic iterations. The skill imports your performance data via MCP / CSV / paste, validates the candidate through the 4-Filter Gate, and refuses to iterate on fake winners.
Mode B — Competitor Research. You don't have a winner yet. The skill chains into Competitor Ad Spy to identify saturated patterns (table stakes, with your brand twist) and uncontested gaps (whitespace), then generates iterations for either lane.
A branded DOCX report with:
- Executive summary and preflight check summary
- Mode and rationale
- Derived account benchmarks (sales-objective only)
- Candidate winner breakdown with 30-day performance against your own benchmarks
- 4-Filter Validation Gate results
- DNA diagnosis — what to preserve across iterations
- Iteration strategy with Andromeda Template Rule rationale
- All 10+ iterations organized by risk level (safe → medium → experimental), each with a complete production spec (hook, shot list, on-screen text overlays, music direction, end card)
- 3-wave testing roadmap with budget thresholds, kill / iterate / scale rules
- Decision framework — what every result teaches you
- Iteration ceiling warning + handoff to PLAN-tier skills (Ad Angle Generator, Pain Point Miner, Competitor Ad Spy) when the concept is exhausted
Edit references/branding.md with your own brand colors, typography, and document structure. The skill reads this file to generate reports that match your brand identity.
Prompt:
Iterate on a winning Meta ad in my [brand] ad account.
Focus on Instagram Reels placement.
Output: Preflight Check confirms Meta Ads MCP is connected, surfaces the user's ad accounts grouped by business, lets them pick the target account. Pulls last 30 days at ad level filtered to sales-objective campaigns. Derives account-specific benchmarks (top-quartile median CTR / CPM / CPC / CPA / hook rate / hold rate) from the account's own data. Identifies the candidate winner, runs it through the 4-Filter Validation Gate (Spend / Signal / Stability / Scale) and surfaces a verdict (Pass / Pass with caveat / Fatigue Warning / Inconclusive). Diagnoses the winning DNA across 7 components and calls out which elements are likely driving performance. Generates 10 Reels-specific iterations (3 safe + 4 medium + 3 experimental), each with a full production spec (hook, shot list, on-screen text overlays, music direction, end card). 3-wave testing roadmap with budget thresholds anchored to the derived benchmarks. Decision framework included — every test outcome teaches you something specific about your audience and pre-commits the next test.
Media buyers, creative strategists, brand owners, and performance marketers who have a winning Meta ad and want to extend its life systematically — without burning the audience out by running it until CPA collapses. Or who want to enter a new category strategically by iterating into competitive whitespace.
This is skill 7 of 10 in the AI Skills for Media Buyers series by Baweja Media.
Built by Baweja Media · MIT License