The Static Ad System

Before the system vs with it

Same product, same reviewer. The only difference is the method now written into these skills. All 162 ads: /all-ads.

The 6 steps

Same steps for every brand. Then repeat from step 3 with the next 10 ads.

1. Brand kit
   Fill the brand's voice, facts, offer, proof, avatar, banned words, product photo. Never invent a fact: unknown = [MISSING], ask.
   Open: BRAND-KIT.md brand.json
2. Understand
   Product → mechanisms → angles. Who buys (avatar). Find ads that win (research). Decode each template: what it says, its visual trick, how it was made.
   Open: 1-understand/
3. Make
   One spec per template (what to KEEP, what to SWAP). The checker blocks bad specs. Generate, paste the real product + exact text by script, check every ad at full size.
   Open: 2-make/
4. Review
   10 ads per batch on a review page. The founder records a Loom. Transcribe it, write each verdict.
   Open: 2-make/review-and-pages.md
5. Improve
   Every comment becomes a permanent rule (and an automatic check when possible). Fix the misses, keep the wins frozen.
   Open: 3-improve/
6. Measure
   First-try rate, after-one-fix rate, final rate, rounds per ad. Compare with the last batch.
   Open: 3-improve/stats.md

The 12 golden rules

Everything else is detail. The full rulebook (with the case each rule came from): 3-improve/rules.md.

1. Clone a winner: same skeleton, new flesh. Change the minimum. The template already works.
2. Read the template as OUR customer would. If a line already works for us, copy it word for word.
3. Copy the visual trick, and how it was made. A photo with a pasted product → edit that photo and swap only the product. Keep the format (9:16 stays 9:16).
4. The product is always the real one. The real cutout and the exact text are pasted by script; an AI never draws the label or the text alone.
5. Every line carries the brand's core problem in plain words (Wildhorn: "beer belly", never just "belly"). Simplest words win.
6. Real facts only. Price, offer, stats, guarantee come from brand.json. Never invent a number.
7. Our customer, not theirs. Our avatar's age and look; remove every prop of their brand.
8. The picture must act out the line. Before/after stays before/after; if she initiates, she leads.
9. Logo outside the label = the icon only. AI misspells brand names.
10. A validated version is frozen. Redo only what was criticized. The founder's latest verdict wins.
11. "OK" means a new file exists on disk. Check the template path, check the output, then QC at full size.
12. Every review comment becomes a rule. Rules are only added, never deleted. The checker blocks repeats.

The folder map

At the root of each skill: SKILL.md (the front page), BRAND-KIT.md + brand.json (the brand). Then:

1-understand, 2-make, 3-improve and scripts are identical in both skills. Only the brand differs.

The review loop

Every review makes the next batch better. Nothing learned is ever lost.

1. Review page
   Only the ads still to review: template next to every try.
2. Loom
   The founder reviews ad by ad, saying the number.
3. Transcript
   Downloaded and transcribed automatically, then read in full.
4. Verdict per ad
   Validated, or rejected with the reason. A validated try is frozen.
5. Why / root cause
   Every win gets its why. Every miss gets its root cause, not the symptom.
6. New rule
   Written into the rulebook with the ad as evidence.
7. Automatic check
   If a machine can check it, the checker now blocks any spec that breaks it.
8. Redo + next 10
   Redo only what was named, plus 10 new templates, built against every rule.

Loom coverage — no comment left without a rule

static-ads

Everything to make perfect static ads. You bring the brand kit.

One skill for the whole job, for any brand: understand the product and the customer, find ads that already win, clone them with the real product, check every image at full size, review 10 at a time by Loom, and turn every comment into a permanent rule. The only input it needs is the brand kit (voice, facts, offer, proof, avatar, product photo).

When it triggers

Just say it, in English or French. For example:

“Make ads for brand X”“Clone these templates”“New batch”“Analyze this ad / competitor”“Find the angles”“Here is my Loom”“Fais-moi des ads”“Trouve les angles”

What's inside — 49 files

wildhorn-static-ads

The same skill, with Wildhorn already loaded.

Exactly the static-ads skill, plus the Wildhorn layer: product facts, the locked ad offer, the angles and headlines that won, banned words, Adrien's taste, all 162 validated ads and every failure. A new Wildhorn batch starts where the last one ended — you only bring the templates.

When it triggers

Just say it, in English or French. For example:

“Make Wildhorn ads”“New batch”“Clone these templates for Wildhorn”“Here is my Loom”“Beer belly ad”“Fais-moi des ads Wildhorn”

What's inside — 58 files

Using the skills

In Claude Code (Mac or VPS)

Already installed. Nothing to set up — just ask. For example: “Make me a batch of 10 statics for brand X from these templates” “New Wildhorn batch from these templates” “Here is my Loom”

On claude.ai

1. Download the zip (buttons above).
2. Settings → Capabilities → Skills → Upload.
3. Upload the zip. Then ask the same way.

For a new brand

1. Brand kit
   Fill BRAND-KIT.md (voice, facts, offer, proof, avatar, product photo). Unknown = [MISSING], never a guess.
2. Understand
   The new-brand worksheet: mechanisms, angles, avatar, templates. You sign off.
3. First batch of 10
   first-batch.md: templates → specs → check → generate → full-size QC → review page.
4. Your Loom
   Review ad by ad. Every remark becomes a rule; the next 10 start from there.

Open points

1. Proof stat wording differs. Landing: “93% felt it*”. Ads: “93% saw their beer belly shrink*”. The compliance audit also flagged 93% vs 84% and 60 days vs 12 weeks. Pick one source of truth.
2. Customer count. Old ads used “20,000+ men”; current is “50,000+”.
3. Autoship. The internal product notes list “Autoship -50%” in the offer. Fine on the site — never in ads (rule).
4. Two reviews were never transcribed: Loom 1 and the final Sam Loom (S11). Their lessons come from the protocols log.