The lessons
A lesson is a technique, explained. No prompt to copy, just the thing that turned out to be true.
Each one is a working rule or a method, written up with what it cost to learn. Where a lesson names a spell you can install, the spell is in the directory.
Watch a check fail once before you trust it to pass
Plant the mistake a check exists to catch, run it and see it go red, then put things back. A search for something that should be absent gets a known hit run beside it, so an empty result can't pass for a clean one.
Undo a bad AI answer instead of correcting it
When a turn goes wrong, rewind to the message before it, change the ask and send it again, so the wrong answer never becomes part of what the AI reads next.
Stop your AI agent from forgetting your project between sessions
Put the project's memory in a few plain text files inside it, each with one job, and the agent reads each one at the moment it's needed.
Treat AI agents like a product team and give them roles
The person who checks the work shouldn't be the one who did it, which is the idea behind the reviewer. The planner, QA, the art director and the producer are the rest of the team, and each has a recipe that does its job.
Write AGENTS.md rules that survive the session
Write down the failure that earned the rule, end it in a check that visibly fails when the mistake comes back, and file it where the agent reads it at the start of every session.
Judge your AI agent's work by using it, not by reading the code
Try every change yourself before your agent saves it, and leave the rest to automatic checks. A report says what was written, and only using the thing shows what it does.
Dictate the thinking and let the AI write the document
Build a template and rules for the format from real examples, then talk the work through and let the AI turn the transcript into the ticket, brief or report your job expects.
Give your AI a map of your files instead of the whole pile
Split the long file your AI keeps rereading into one small file per item, and turn the original file into a lean index, one line per item pointing at each one's file, so the AI opens only the item the task needs.
Stop at half the context window
Close the session yourself and start again from a note you can read, before the tool summarizes the conversation on its own and you carry on from a summary nobody checked.
Teach the AI with examples a person actually wrote
Asked cold, a model guesses at average. Give it real examples of the output done well and the best guide you can find, and check a person wrote each example, because rules learned from machine-written examples copy the machine.