Paths, pages and commands — the Wildhorn static-ad workshop

Everything runs on the VPS (ssh vps-claude). Never run Playwright/Chromium or generation on the Mac. Secrets are referenced by path only.

---

1. Workshop layout — /root/workspace/wh-remix/

2. Sam-batch workshop — /root/workspace/wh-remix/sam-batch/ (the current pipeline)

3. Published pages (all noindex, clean URLs)

Static root: /srv/blueprint/ (ex/ = full-size tries, ex/web/ = webp thumbs, all-ads/b1 b2, wildhorn-ads/).

4. Mac exports (~/Downloads/)

Export commands (run on the Mac — the only thing allowed locally):

rsync -a vps-claude:/root/workspace/wh-remix/sam-batch/sell/ ~/Downloads/Wildhorn-Sam-Batch/
rsync -a vps-claude:/root/workspace/wh-remix/sam-batch/validated/ ~/Downloads/Wildhorn-Validated/
rsync -a vps-claude:/root/workspace/wh-remix/run1-final/ ~/Downloads/Wildhorn-Final-Ads/

For a new batch, create a new folder (e.g. ~/Downloads/Wildhorn-Batch-3/), never overwrite the previous exports.

Templates from the Mac to the VPS (Adrien drops them in iCloud BUSINESS/WILDHORN CONCEPTS/, a zip in Downloads, or the TEMPLATE WINNER folder ~/Library/Mobile Documents/com~apple~CloudDocs/BUSINESS/E-COMMERCE/- ADS ECOM/TEMPLATE WINNER/):

scp -r "<local folder>"/* vps-claude:/root/workspace/wh-remix/<new-batch>/templates/

5. Memory & checkpoints

• Checkpoint (read at the start of any Wildhorn batch work): ~/.claude/bin/checkpoint wildhorn-batch-test (stored on the VPS in /root/workspace/_checkpoints/). Write: ~/.claude/bin/checkpoint wildhorn-batch-test --write (stdin).
• Mac memory notes: wildhorn-remix-system.md, creas-gpt-image-2.md, clone-winner-system.md, english-deliverables.md, wildhorn-lp-compliance.md (project memory folder of the Mybusiness project).
• Vault copy of the rulebook: /root/claude-vault/knowledge/copywriting/TEMPLATE-CLONE-LESSONS.md.
• Research: /root/claude-vault/knowledge/product-research/primal-bison/etude-marche-primal-edge.md, structured /srv/intel/data/primaledge.json.
• Landing (product claims source): /srv/wildhorn-lp/product-v6-monument.html.

6. Command cookbook

Lint a spec file (always first)

ssh vps-claude 'cd /root/workspace/wh-remix && python3 lint.py sam-batch/s5specs.py'

Template-path sanity check (D13 — not in lint yet)

ssh vps-claude 'cd /root/workspace/wh-remix/sam-batch && python3 - <<EOF
import importlib.util, os
T = sorted(os.listdir("templates"))
sp = importlib.util.spec_from_file_location("m", "s5specs.py"); m = importlib.util.module_from_spec(sp); sp.loader.exec_module(m)
for s in m.S:
    if not s.get("tpl_abs") and s["tpl"] != T[s["num"]-1]: print("MISMATCH", s["num"], s["tpl"], "->", T[s["num"]-1])
print("checked", len(m.S))
EOF'

Generate one ad

ssh vps-claude 'cd /root/workspace/wh-remix/sam-batch && python3 srun.py s5specs.py 84'

Output: #84: openai OK (or REFUSED -> seedream_comp then seedream+comp OK). D10: OK only if a new out/s_84_vK.png exists.

Generate a round in parallel (10 ads, 5 at a time), with logs and a done flag

ssh vps-claude 'cd /root/workspace/wh-remix/sam-batch && echo "84 85 86 87 88 89 90 91 92 93" | tr " " "\n" > jobs5.txt && \
nohup sh -c "cat jobs5.txt | xargs -P 5 -I{} sh -c \"python3 srun.py s5specs.py {} > logs/r5_{}.log 2>&1\"; touch logs/_r5done" >/dev/null 2>&1 &'

Check: ssh vps-claude 'ls /root/workspace/wh-remix/sam-batch/logs/_r5done; tail -n1 /root/workspace/wh-remix/sam-batch/logs/r5_*.log'. ~110 s per gpt-image-2 image.

Manual fallback (sexual / refused) — Seedream scene + composite

# spec has engine="seedream_comp" (or fallback="seedream_comp"), sd_prompt (scene only: no text, no product, empty zones), composite {...}
python3 srun.py s5specs.py 88
# by hand:
python3 ../kie_gen.py s_88_sd --prompt prompts/s_88_sd.txt --product templates/<tpl> --model seedream [--aspect portrait_16_9]
python3 ../composite.py ../out/s_88_sd_v1.png ../out/s_88_v2.png prompts/s_88_comp.json

Painted-in text (sunscreen, chalk…) — Nano Banana Pro

python3 ../kie_gen.py s_88_nano --prompt prompts/s_88_nano.txt --product ../product-jar.jpg --concept templates/<tpl> --model nano

A25 — copy THEIR production method (minimal edit of the template)

# 1) Seedream edits the template itself: age the man, clear the text, nothing else (keep aspect)
python3 ../kie_gen.py s_11_sd --prompt prompts/s_11_sd7.txt --product templates/09-14_c0025_Primal-Viking.png --model seedream --aspect portrait_16_9
# 2) swap only their jar for ours inside a box (pixels of the base image)
python3 ../jarswap.py ../out/s_11_sd_v6.png prompts/s_11_jar.png X0 Y0 X1 Y1
# 3) exact text by script
python3 ../composite.py prompts/s_11_jar.png ../out/s_11_v5.png prompts/s_11_comp5.json

"Change only the wording" on a liked try (A22)

Spec with tpl_abs=O + "s_29_v2.png" (O = /root/workspace/wh-remix/out/) → Image 2 = our own try.

Record verdicts after a Loom

ssh vps-claude 'cd /root/workspace/wh-remix/sam-batch && python3 - <<EOF
import json
v = json.load(open("validated.json"))
v.update({"84": "✅ Validated (first try)", "85": "❌ Loom 6: off-angle (A24)"})
json.dump(v, open("validated.json", "w"), ensure_ascii=False, indent=1)
so = json.load(open("sell_override.json")); so.update({"86": 1}); json.dump(so, open("sell_override.json", "w"))
EOF'

Rebuild every page + exports

ssh vps-claude 'cd /root/workspace/wh-remix && python3 sam-batch/build_sam.py && python3 sam-batch/build_samfinal.py && python3 sam-batch/build_validated.py && python3 build_all.py && python3 build_lessons.py'
rsync -a vps-claude:/root/workspace/wh-remix/sam-batch/sell/ ~/Downloads/Wildhorn-Sam-Batch/

(For a new batch folder, copy the builders and change their paths/titles; keep SKIP_VALIDATED on the review page.)

Transcribe Adrien's Loom

ssh vps-claude 'cd /root/workspace/wh-remix/feedback && mr-transcribe loom-s5.mp4 --lang en'

(download the Loom mp4 to that folder first; API transcription only.)

Hit-rate numbers

ssh vps-claude 'cd /root/workspace/wh-remix/sam-batch && python3 -c "
import json; v=json.load(open(\"validated.json\")); c=[json.loads(l) for l in open(\"calls.jsonl\")]
ft=sum(\"first try\" in x for x in v.values()); ok=sum(x.startswith(\"✅\") for x in v.values())
print(f\"first try {ft}/{len(v)} = {100*ft/len(v):.1f}%  validated {ok}/{len(v)}  calls {len(c)}  per validated {len(c)/max(ok,1):.2f}\")"'

7. Reference numbers