Blog
Field notes from running AI assistants on real documentation: what works, what breaks, and what we changed.
-
Why your AI assistant needs more than your text
An AI assistant never reads your page. It reads a passage cut out of it, plus whatever signals travelled along: the title, the section, the version, the date. Lose those and the best-written content gives wrong answers. Here is what that looks like, and whose job it is to fix.
-
How to close the documentation feedback loop with AI: from a thumbs-down to a fixed page in one chat
Most docs teams collect feedback and stop there. Here is a workflow that takes a reader's thumbs-down all the way to an updated page, re-read by your docs assistant and marked done, inside Claude, ChatGPT or Claude Code.
-
Cheap models fail quietly: small vs large LLMs on real documentation
We ran Mistral Small 4, Mistral Large 4, GPT-5 Mini and GPT-5.2 on the same documentation passages. Here is where the small models break, and what it costs to avoid it.
-
How to make your content AI-ready
Your documentation can be easy for people to read and still useless to an AI. Here is where the meaning gets lost between your pages and the answer, ten rules for content that survives the trip, and a ten-minute check for your own site.
-
Why most docs chatbots can't read your documentation
Modern documentation sites are web apps: the text your visitors see is not always in the page a simple crawler downloads. Here is how that breaks an AI assistant, how to check your own site in five minutes, and what reading a site properly involves.