The judgment layer of a production Google Ads agent — operator-grade diagnosis frameworks, LLM self-check gates, and a closed loop with read-back verification. Checklists audit accounts; judgment runs them.
SaferSkills independently audited ads-agent-brain (Agent Skill) and scored it 100/100 (green). The audit ran 55 deterministic rules across Security, Supply Chain, Maintenance, Transparency, and Community; it found 0 high-severity and 0 lower-severity findings. The full rule-by-rule trace and per-finding evidence are below. Free, methodology-open.
Findings & checks · 0 flagged
Every scanned point with the score it earned and what moved between them.
First recorded scan — no prior version to compare against.
The primary manifest — the file an agent reads to learn what this artifact does.
Checklists audit accounts. Judgment runs them.
This repo is the cognitive layer of a production Google Ads agent — the distilled judgment of a strategist with 5+ years running cross-border e-commerce accounts, encoded into files an LLM can actually follow. It has been running daily, autonomously, on real accounts with real budgets since March 2026.
Most open-source "AI + ads" projects are audit checklists: 250 checks, a weighted score, a PDF of recommendations. Useful once a quarter. This is the other thing — the part nobody ships: how an operator actually thinks when money is moving, written so that an agent can run the loop every single day:
pull everything → diagnose → generate candidates → tiered authorization
→ execute → read-back verify → record → review outcomesThree things in here that we haven't seen anywhere else:
This is a distilled snapshot, not a dump of our production system:
<CLICK_THRESHOLD>, <STEP_SIZE>, …). The numbers we run were tuned over months against specific accounts; yours differ. We ship the judgment — you calibrate the numbers. docs/calibration.md tells you how, in about two weeks of data.| Path | What it is |
|---|---|
| brain/BRAIN.md | The cognitive layer: first principles, domain knowledge, thinking frameworks, diagnosis framework, LLM self-checks, operator heuristics |
| skills/daily-ops.md | The daily execution manual: full GAQL pull set, candidate pipelines (exclude / enable / add keyword / add negative), expansion & contraction evaluation, briefing contract, execution protocol |
| skills/QUERIES.md | GAQL query catalog — every query battle-tested against API v23 |
| skills/EXECUTION.md | Mutation protocol: verify-before-mutate, read-back verification, operation journaling schema |
| docs/architecture.md | How the pieces form a closed loop, and why each gate exists |
| docs/calibration.md | How to turn every placeholder into your number |
| zh/ | 中文版（原生版本——this system was authored in Chinese first） |
brain/BRAIN.md into your agent's always-on context (Claude Code: reference it from CLAUDE.md).skills/daily-ops.md as a skill / slash command.<PLACEHOLDER> with numbers calibrated to your account.These principles are themselves documented in the brain — the repo demonstrates the method it describes.
MIT. Use it, fork it, run your accounts with it. If it makes you money, a star is appreciated; a story in Issues about what it caught (or missed) is appreciated more.
~30 seconds. Free. No account. Every finding cites a rule and a line of evidence.