Your documentation always had one reader.‸Now it has a second one.

Why your AI gets it wrong

When a Copilot or Claude rollout answers confidently and incorrectly about your own systems, it is rarely the model. It is what the model was given to read. Your systems were documented for people — people who could ask a coworker when something didn't make sense. The AI can't ask. So it guesses.

In 21,000+ controlled tests, how an API was documented explained over 10× more of the variance in AI‑generated code quality than which model was used. Better docs beat a better model. That research is published, and it is the method behind everything below.

How it works

1. ShowA free one-hour session with your team. We put your real documentation in front of the AI, give it real tasks, and watch where it breaks — together, on screen.
2. AuditThree tasks your people actually do. Two weeks. Written findings, a prioritized fix list, and a baseline measurement of what the AI gets right today.
3. Fix & measureRemediate at the source, connect the AI to it properly, and re-run the same tasks. You get the before-and-after number, not a vibe.

Every finding is reproduced live or labeled unverified. No exceptions.

Who's behind it

Ed Grzetich — twenty years as a technical writer across General Dynamics (including four years leading documentation on a NASA program), Mastercard, and AWS, where he owned API and SDK reference documentation.

• Tokens Not Jokin’ — the research: 21,000+ tests on docs & AI code quality
• artie-cli — open-source OpenAPI scoring tool, on PyPI
• grzeti.ch — tools, writing, and portfolio

Co-leader of Fort Wayne AI, one of Indiana's most active AI practitioner communities.