Built for operators running $5K+/mo on Google Ads who want agentic loops on the account — without the Monday-morning cleanup bill.
Install across 4 clients
Claude Desktop, Claude Code, Gemini CLI, and ChatGPT (Developer Mode). The exact pipx + ADC setup and the auth gotcha that bites first-time installs.
12 GAQL prompt templates
Wasted spend hunt, brand cannibalization scan, asset decay, AIO impression decay, auto-apply lockdown, anomaly detector - all with actual GAQL.
CLAUDE.md governance template
100+ lines: account context, brand-safe terms, never-block words, hallucination guardrails, approval gates.
Claude Skills config
The .claude/skills/google-ads/ layout - SKILL.md, reference docs, prompts, scripts. Drop it in once; every future session loads with account context.
Read-only vs write-capable comparison
Official Google Ads MCP (3 tools, read-only) vs cohnen, GoMarble, Markifact, and AdKit - which add write access, and what to audit before you trust them.
Kill-switch policy
Spend velocity 25% deviation trigger, write approval gates, audit log requirements, the daily 50-write ceiling.
No verification step in negative-keyword generation
Claude happily generates 200 plausible negatives without running a single GAQL. The fix is a 4-line CLAUDE.md policy that requires verification on every proposal - included.
Conversion definitions never make it into context
The agent doesn’t know what counts as a primary conversion vs Journey-Aware Bidding secondary. Without it, every recommendation drifts. The 20-line block fixes it.
No audit log when the agent writes
The append-only .jsonl log pattern - what your client, your board, and your legal team will eventually ask for.
“The CLAUDE.md template is the actual artifact. Every agency stealing the rest of this content is missing that one file - and that's the one that prevents the 200-hallucinated-negatives failure.”
“The wasted-spend GAQL prompt found $4,200/mo in week one. Ran it on three more retainer accounts. $11.8K total recovered in 30 days.”
“I was running a premium video model on every ad, and the per-output cost was brutal. The model-by-job matrix told me to use Kling for the bulk and save Veo for the hero spots. Cost dropped 78% the next month.”
“Showed the AI Slop Detector to our designer. We scored five live ads, three came back at 6+ flags. That's the 18% CPM tax we'd been paying without knowing.”
“Set up the Skills folder in 40 minutes. Every Claude session now loads with our Klaviyo context. Eliminated the 'paste in the voice doc' step every prompt used to need.”
“New hires get this on day one. The 12 GAQL prompts compress 6 months of 'how do you write that query again?' into one folder they can read on a flight.”