The split from this week’s issue: AI is a historian, great at mapping what already worked. Humans make the forward bet, the angle with no data yet. Then AI scales it. Two paste-ready prompts below, one for each half. Copy the blocks, fill in the [BRACKETS], paste into Claude Code.
Wire your Meta account into Claude, pull every ad you have ever run, and tag each one by product, persona, angle, and emotion against ROAS and CPA. You get a map of what is proven and a map of the white space you have never tested.
Create a Meta developer app at developers.facebook.com, add the Marketing API plus ad-performance scopes, copy the App ID and App Secret, create a System User and assign it to your ad accounts and pages, then generate a token scoped to ads_read (it may need a second admin to approve, up to an hour). Grab each ad account ID (act_XXXXXXXXXX). The full step-by-step is in the downloadable kit.
META_APP_ID=[your app id]
META_APP_SECRET=[your app secret]
META_ACCESS_TOKEN=[your system user token]
META_AD_ACCOUNTS=[act_xxx:US, act_yyy:DE, act_zzz:FR]
META_ROAS_TARGET=[e.g. 1.2]
META_CPA_TARGET=[e.g. 35]
ANTHROPIC_API_KEY=[your anthropic key]
GEMINI_API_KEY=[your gemini key]
You are my creative-history analyst. Use the Meta Marketing API credentials in .env to pull and tag every ad I have ever run, then grade them.ACCOUNTS: read META_AD_ACCOUNTS from .env. Treat each id:label pair as a separate market.TARGETS: read META_ROAS_TARGET and META_CPA_TARGET from .env. These define winner vs. loser.For each ad account:1. PULL. Fetch every ad with: ad name, status, creative (image/video thumbnail URL), primary text/headline, spend, purchases, revenue, ROAS, CPA, CPM, date range. Page through the FULL history, not just the last 30 days.2. READ THE CREATIVE. For each ad, use the Gemini key in .env to look at the actual image and read what is on it: product shown, on-image copy, format (UGC, static, testimonial, offer, meme, comparison).3. TAG each ad with: product, persona (specific, NOT a vague label), angle, emotion, format. If a tag would come out vague (fitness enthusiast, health-conscious), flag it UNDERSPECIFIED. Do not invent specificity.4. GRADE against the targets. Sort every tag combination into SCALING / HOLDING / WASTE. Show the spend behind each so I know what is carrying the account.5. OUTPUT two markdown files per market:- proven-map-[label].md : ranked tables of persona x angle x emotion x format by ROAS, with spend.- white-space-[label].md : personas, angles, formats, products, and occasions I have NEVER run, or run with trivial spend. Be exhaustive and honest.Stop and ask me before pulling if any account returns zero ads (token/scope issue). Re-runs: only pull ads newer than the last run, then merge.
Have the actual research session: two people, Reddit, TikTok, and your reviews open, riffing on angles you have never run. Record it. Then drop the transcript in and let the machine turn the mess into briefs scaled to your weekly count.