AdLoop

A self-learning ad creative engine. Give it a brand URL; it researches the brand, writes art-direction briefs where every claim traces to a sourced fact, renders finished ads (gpt-image-2), verifies its own output by reading the pixels back, and learns from every batch and every founder edit — so batch N+1 is measurably better-behaved than batch N.

Built at Push to Prod (Bengaluru, 2026-08-08) on Claude Managed Agents as the runtime. Submission: https://devfolio.co/projects/adloop-7e59

The self-learning loop — with receipts in this repo

The core claim is that the engine learns. Every step of that claim is verifiable in tracked files, from one live run on godesi.in (an Indian D2C candy brand):

1. Batch 2 shipped; the founder made two edits asking for "festival mela energy" → brands/godesi.in/batches/002/edit-ad-1/ and edit-ad-2/ (the thin edit loop: smallest brief change, single-ad re-render, re-verify).
2. Two edits with the same lesson = a permanent rule. The engine generalized both edits into one promoted rule — written by the agent, unprompted, into brands/godesi.in/memory.md: "make it festive / like a mela" = add triangular-flag bunting, change nothing else — including which brand color the flags take per ground, and a caution not to violate the design guide's one-magenta-element cap.
3. Batch 3 applied the rule with zero new instructions. All three briefs in brands/godesi.in/batches/003/briefs.json carry a mela-flag line in their layout slot. Nobody asked. That is the loop closing.
4. Repetition is a decision, never an accident. brands/godesi.in/history.json is the ledger: every shipped ad keyed by angle × SKU × persona with a status (shipped → approved/edited/killed → winning/fatigued). Batch 3's mix rationale (brands/godesi.in/batches/003/plan.json) shows winners re-run deliberately with new scenes, unproven cells never duplicated, and one exploration slot always kept.

The truth layer — it rejects its own ads

Every fact is captured with a source URL into a number bank with an explicit forbidden list (brands/godesi.in/inputs/facts.json). Every brief carries facts_used[]. After rendering, a vision pass reads the finished image and diffs every numeral on canvas against the facts:

• brands/godesi.in/batches/003/verdicts.json: all three batch-3 ads FAILED verification — the image model hallucinated numerals into the scene (sticky notes reading "Client call 3:30 PM", "Q3 Review 4:00 PM"; an untraceable ₹5 badge) and the verify pass caught every one. Refusing to ship those IS the product working.
• Batches 1–2: zero untraceable numerals across six ads (brands/godesi.in/batches/002/verdicts.json).
• Validated before the build: 9 ads across 3 real businesses, zero invented claims — while a captured competitor pipeline invented guarantees, review counts and discount codes (docs/COMPETITOR_TEARDOWN_MAKELOCALADS.md).

Architecture (Claude Managed Agents)

Chat / canvas ⇄ host CLI (engine/)
  ├─ capture  — 1 CMA session, multiagent: 3 parallel research lanes (facts, design, personas);
  │             input QA gate (view every ref image) is never delegated
  ├─ batch    — 1 CMA session: read ledger → decide mix → write briefs (pinned 8-slot schema)
  │             → assemble → render (KIE gpt-image-2, per-ad refs) → vision verify → ledger append
  ├─ edit     — thin loop: smallest brief change → single-ad re-render → learn (edit log → rule promotion)
  └─ memory store (per brand) = ledger + learned rules + fact bank, mounted read-write into every session

• Vault credentials: the render API key is substituted at egress — the model never sees it.
• Budgets: per-session hard caps with pause/resume (budget_reached pauses, host raises, work continues).
• Pinned schemas: every cross-agent artifact shape is validated host-side (engine/schemas.py); parsers accept only the pinned shape.
• Stage runbooks (the full orchestration prompts): engine/prompts.py.
• Runtime decision + spike evidence: docs/SPIKE_RESULT_CMA_GO_2026-08-08.md.

Run it

python3 -m venv .venv && .venv/bin/pip install anthropic     # + .env.local with keys
python3 -m engine setup                                      # standing cloud resources
python3 -m engine capture <url>                              # cold-start a brand
python3 -m engine batch <domain>                             # one batch, end to end
python3 -m engine edit <domain> ad-2 "make it more festive"  # founder edit → re-render + learn
python3 -m engine canvas <domain>                            # live canvas at localhost:8787

Canvas pages: / (ad cards, verdicts, ledger, rules) · /start?replay (the six pipeline stages replayed with the latest batch's real artifacts) · /memory (edits → rules → applied unprompted).

Where to verify each claim

Claim	File
Founder edits become rules (2 = promote)	brands/godesi.in/memory.md, batches/002/edit-*/
Next batch applies rules unprompted	brands/godesi.in/batches/003/briefs.json (layout slots)
Verify pass rejects hallucinated numerals	brands/godesi.in/batches/003/verdicts.json
Ledger: repetition is a decision	brands/godesi.in/history.json
Truth layer: number bank + forbidden list	brands/godesi.in/inputs/facts.json
Pinned cross-agent schemas	engine/schemas.py
Full stage runbooks (orchestration)	engine/prompts.py
Ground truth the build reproduces	reference/scratch-tests/ (3 validated end-to-end runs)
How the architecture was derived	docs/ORIGINS.md, CLAUDE.md

Rendered ad images are not tracked (size); see the Devfolio gallery or run the canvas.