Plan: brand in, winning ads out (automatic)

Goal. Load a brand and the system does the rest: it finds today's winning ads, picks its sources by a written rule, and generates finished ads for review.

A. Build the machine (once, for every brand)

  1. One command for stage 1: find-winners <brand>. It runs two lanes.
    • Native (top of funnel). ScrapeCreators pulls Meta's own impression order for the buyer's pain and product words. It keeps ads launched 30–100 days ago, one per advertiser.
    • Middle and bottom funnel. GetHookd pulls designed ad styles: 5★, top 10% of the ad's own brand, 30–100 days, landing on a product page, one per brand.
    • The command also downloads every image, builds contact sheets, and checks every ad is live on Meta by exact ID.
  2. Image review (done by the agent, written to a file). For each ad:
    • Does it sell to the buyer's pain?
    • What funnel stage is it, judged from the image? Nothing on it = top, hook or claim = middle, price or offer = bottom.
    • Is the whole pitch on the image?
    • What format is it?
  3. Automatic pick of 3 sources, in this order:
    • relevance
    • pitch on the image (pain or claim plus offer)
    • evidence: 5★, live, advertiser size
    • one per format
    • skip any source already used
  4. Write it into the skills.
    • find-printing-ads: stage 1 plus the pick rule.
    • make-printing-format-statics: takes the picked sources automatically.

B. Run CHL end to end (the first test)

  1. Stage 1. Run it fresh for CHL with the one command. Cost: about 5 GetHookd and about 16 ScrapeCreators credits.
  2. Pick. Auto-pick 3 sources. Today's preview: mindbodygreen, Healthline × Ritual, EllaOla "MTHFR Gene vs With EllaOla".
  3. Concept cards. Each concept copies its source's layout and uses CHL's own pain-first words. CHL's rules apply:
    • FIRST BOTTLE FREE is the second-biggest element
    • no shipping mentioned
    • real proof only: "4.9 · 2,600 five-star reviews"
    • capsules at real scale
    • the link goes straight to the product page
  4. Primary text. Borrow the native winners' structure: a first-person story in the caption. That is where their selling lives.
  5. Prompts. Write them, then run the pre-generation gates: source lock and single image.
  6. Generate with Codex image_gen and the real product references, 4:5 first.
    • Version B: the whole ad in one pass.
    • Version A: a clean photo with hand-set type at B's weight.
  7. QA. Check the label at 100%, legibility, type kept off the product, distinct layouts, and the phone-size view. Any failure means a new prompt and a fresh generation. Never edit an image.
  8. Review page. A light page (≤1 MB) with each source beside our ad.
  9. Launch. You approve what launches. It uploads to Meta paused and goes live on your OK.
  10. Results. Read them against your kill line.
    • Winners: make more from the same source family.
    • Losers: pull new sources.
    • Log everything to the ledger and history.

C. Then

  1. Run the same command for RAW and Wonder: 3 per brand first, then 10 once it holds.

Budget and risks