First batch — from a filled brand kit to the first 10 ads

Step by step for a new brand, once BRAND-KIT.md is filled (and, for a brand with no research yet, once 1-understand/new-brand-worksheet.md is signed off). The product images, brand.json field by field and what is Wildhorn-only are in BRAND-KIT.md. The full pipeline behind each step: 2-make/PIPELINE.md.

0. Division of labour inside this skill

• 1-understand/ does the research and the reading: product sheet, mechanism map, competitor ads, avatar, top-5 angles + headline bank, template intake + concept families, and one hand-off brief per template (1-understand/new-brand-worksheet.md §9). Run it first, with its two sign-off gates (angles/avatar; first concept family of briefs). Don't duplicate it here.
• 2-make/ turns briefs into specs, generates, QCs and builds the review pages; 3-improve/ runs the review loop and keeps the rulebook.
• If the reviewer hands you templates without a brief, do the decode yourself in the spec (why, dna, scene, inventory) using 1-understand/template-reading.md as the checklist.

1. Working folder

ssh vps-claude
mkdir -p /root/workspace/<brand>-remix/{batch/templates,batch/prompts,out,feedback,pages}
cd /root/workspace/<brand>-remix
cp /root/claude-vault/skills/static-ads/brand.json brand.json
export SAC_WORKDIR=$PWD SKILL=/root/claude-vault/skills/static-ads

Everything a batch produces stays in this folder; never write into another brand's workspace or into /srv directly.

2. Spec constants for the brand

Rewrite the shared strings at the top of the first spec file (2-make/spec-format.md): OLD (avatar look), the ingredient constant (ORGANS for Wildhorn), the capsule constant, HORN (icon-only sentence), STD (generic KEEP lines). These carry the latest wording of C2/A11/D8 into every ad.

3. Templates

• The reviewer (or 1-understand/new-brand-worksheet.md §8) provides the winning templates. Accept every template (R1); sort into concept families (R2).
• Copy them (never move) into $SAC_BATCH/templates/, named so that sorted order = ad order (MM-DD_cNNNN_Source.ext). The extension must be the real format (D10).
• Read each template at full resolution before writing its spec. Note its aspect ratio and production method (photo, illustration, text-only, photo + pasted product → A25).

4. First batch of 10

1. Briefs → draft specs: python3 $SKILL/scripts/brief2spec.py $SAC_BATCH/s1specs.py <kit>/briefs/*.json (or write specs by hand).
2. Do the judgment by hand: inventory ≥ 5 lines, angle per template, texts slot by slot, visual rewritten as one scene, fallback recipe for sexual/skin ads.
3. python3 $SKILL/scripts/lint.py $SAC_BATCH/s1specs.py → LINT OK.
4. Generate: for n in $(seq 1 10); do python3 $SKILL/scripts/srun.py $SAC_BATCH/s1specs.py $n; done (or xargs -P 5).
5. Full-res QC of the 10 (2-make/qc-checklist.md); fix before showing (QC log).
6. python3 $SKILL/scripts/build_sam.py → wire $SAC_PAGES to a <brand>-review.machineroom.app handle once → send the link.
7. Loom → 3-improve/feedback-loop.md.

5. What to expect

• The generic rules transfer (decode, inventory, text count, placement, engines, QC). Expect the first misses on brand taste: the reviewer's idea of aggressive, the avatar's real behaviours, which angle is default, how the ingredient must look. Those become new rules in round 1–2.
• Expect 40–60 % first try on the first 10, rising to 70 %+ by round 3 if every miss becomes a rule (Wildhorn: 60 % → 76 % → 72 % with 13 QC catches).
• Expect OpenAI refusals on anything sexual or with bare skin close-ups: plan seedream_comp for those templates at spec time.
• Expect the logo to be misspelled whenever it appears outside the label: icon-only from day one (D8).

6. Pre-flight checklist (before the first generation of a new brand)

• [ ] 1-understand/ sign-off gate 1 passed (angles + avatar) and gate 2 for the first concept family of briefs.
• [ ] BRAND-KIT.md and brand.json filled (no [MISSING] left); python3 -c "import json;json.load(open('brand.json'))" OK; lint.py runs without a JSON error.
• [ ] product_photo opens, label readable at full resolution; product_cutout has a clean alpha edge (render one composite on white and on black to check halos).
• [ ] product_lock names every visible packaging element; lock_extra carries the icon-only sentence; icon_description matches the real icon.
• [ ] prices_allowed, offer, proof match the live offer exactly (reconfirm the price with the founder — Wildhorn's was locked at $49.95 only after asking).
• [ ] banned includes competitor names, jargon, compliance bans; angle_keyword set for the default angle; visual_checks cover the ingredient look.
• [ ] Spec constants written (avatar look, ingredient look, capsule look, icon sentence, STD).
• [ ] Templates copied into batch/templates/, sorted order = ad order, real extensions; aspect ratio and production method noted per template.
• [ ] SAC_PAGES served on its own *.machineroom.app handle (noindex, clean URLs) — or the founder knows where to look.
• [ ] calls.jsonl, validated.json, sell_override.json start empty for the batch; rounds.json only if needed.
• [ ] Budget agreed for the round (10 ads ≈ 10–20 OpenAI calls ≈ $3–6 at ~$0.30/image, plus a few Seedream scenes).