Web search hands an AI assistant stale forum posts and deprecated APIs. Shallow submodules of the Home Assistant and HACS docs, symlinked under docs/ and described in CLAUDE.md, give it the real thing.

An AI assistant showed me C#’s $$ raw interpolated strings while I was getting something else done. Learning now happens sideways, and knowing a feature exists doesn’t replace judgement about when to use it.

Have the AI write two files: a short wrapper you read and approve, and a Deno script that runs with only the permissions the wrapper grants. What that protects against, what it doesn’t, and why --allow-run breaks it.

Telling a model not to do something puts that thing in its context, and attention has no way to subtract it. Describe the output you want, not the one you don’t.

A starting template for the instructions file an AI coding assistant reads before it touches your repository. It gets what you know about boundaries and shared code out of your head and into a file the tool reads.

AI coding assistants are too helpful: the code works, but it doesn’t belong. SOLID principles, used as constraints on the assistant, give it the pause an experienced developer has before changing things.