45 Tracking Problems Solved
Claude + Google Tag Manager — read the container, write the spec, build in a workspace, publish only when you say so
There is a post going around saying Claude plus Google Tag Manager means you can stop doing tracking work by hand. The claim is broadly right, and the detail matters enormously — because what you are being invited to do is give an AI write access to the system that measures your revenue.
This gives you the forty-five prompts. It also tells you how the connection is actually made, what it can and cannot do, and where the human has to stay in the sequence. Google publishes no official Tag Manager MCP server, so the connection runs through a third-party one authenticating as you — which means it inherits whatever permissions your Google account holds.
Every prompt follows the same order: read the live container, draft in a workspace, package a version, QA in preview, and publish only on an explicit decision. That sequence is not decoration. It is what keeps a mistake recoverable.
The four rules, which matter more than the prompts
Read before you write. Every audit prompt ends with an instruction not to create, update, delete, version or publish. If Claude misidentifies the container, a read-only instruction turns a disaster into a wasted minute.
Never publish from a prompt. Claude drafts in a workspace and stops. You review the diff, run the QA checklist, and publish. That line does not move, however confident the output sounds. Record the live version number before you start — that is your rollback point, and it needs writing down before the first change rather than after something breaks.
Treat deletions as a separate, slower decision. A tag nobody recognises is not necessarily a dead tag. Claude is genuinely good at reading two hundred tags and telling you which fourteen matter; it cannot know the strange one was added for a legal requirement in 2023. That context lives with you, which is why the human stays in the sequence.
Read it here
The whole document, in the page. Nothing to fill in, and no need to leave the site — though you can download it if you prefer.
1 / 33
Free to read, quote and share with attribution. Not for resale.
What's inside
Eight clusters, forty-five prompts, each written so nothing reaches the live container without your approval.
- 01
Cluster 1 — Container Audit & Cleanup
7 prompts. Reading the live version to find what is actually running, not what someone intended.
- 02
Cluster 2 — GA4 Foundations
6 prompts. Getting the base configuration right before anything is built on it.
- 03
Cluster 3 — Ecommerce & Purchase Tracking
7 prompts. Purchase events, dataLayer specs and revenue accuracy.
- 04
Cluster 4 — Lead Generation & Form Tracking
5 prompts. Form submissions that fire once, on a real success event.
- 05
Cluster 5 — Google Ads Conversions
5 prompts. Conversion actions, counting settings and enhanced conversions.
- 06
Cluster 6 — Meta, LinkedIn & Other Platforms
5 prompts. The pixels that accumulate in a container and rarely get audited.
- 07
Cluster 7 — Consent, Privacy & Server-Side
5 prompts. Consent Mode, and moving measurement server-side without breaking it.
- 08
Cluster 8 — Debugging, QA & Governance
5 prompts. Preview-mode checklists, version diffs, and explaining what changed when something broke last Tuesday.
Who it's for
- Anyone connecting Claude to a live GTM container
- In-house marketers who own their own tracking
- Developers implementing a dataLayer specification
- Agencies auditing a container they inherited
Free · no email required
33 pages, 392 KB. Free to read, quote and share with attribution. Not for resale.
Download the PDFI'd rather you argue with it than agree quietly
If something in here is wrong, or you have run these checks on your own account and found something different, I want to hear it. And if you want the eighth pillar run against your current agency's reporting, that is a conversation worth having.