Govern: One way of working for your team's AI agents

What you will be able to do

• Replace a long instructions file with a short AGENTS.md that routes and never holds
• Lock the list of standards and docs before the work, review against it, and keep docs current in the same commit
• Block a rule that must never break with a hook, tested both ways before the team gets it
• Ship every bug fix one way: a red test first and a separate reviewer before the commit
• Turn each fix into a team rule, catch what nobody predicted, and cap what grows
• Make every skipped step visible: written where the team reads it, with a named reader
• Govern your own repo, checked against the course rubric

1. 1Make every agent read one fileCheck exactly what each agent loads with /context, then turn a long shared file into one short AGENTS.md that every machine loads. VideoFree preview
2. 2Review against a locked standards listHave the agent write the standards it will follow before it changes code, lock that list in the pull request template, and review every change against it rule by rule. VideoLocked
3. 3Keep docs current in every commitGive one area its own doc with the facts the code cannot tell the agent, name it in AGENTS.md, and add one rule so the same commit updates the doc when the code changes it. VideoLocked
4. 4Block what must never shipAsk for most rules and enforce the few that must never break: a shared Claude Code hook that blocks a billing commit with no test, tested both ways before the team gets it. VideoLocked
5. 5Ship every bug fix one wayWrite the team's bug-fix steps in one workflow file, red first on the value, and hand every fix to a read-only reviewer agent that sees the red itself and runs the cases the docs describe. No commit lands before a review passes. VideoLocked
6. 6Turn each fix into a team ruleAdd one question to every fix: is this a one-off, or a class of mistakes? A one-off stops at the fix. A class gets an edit in its home (a standard, a doc, a workflow, a hook), proposed by the agent and made by the team. AGENTS.md only for a rule every change must follow. VideoLocked
7. 7Catch misses, cap what growsAfter the work, every mistake that reached main after a review passed it becomes one entry in a short failure list the review reads first. Once a month, AGENTS.md, each standard and the failure list are capped, and a count of what the reviews cite tells the team which rule to fix or retire before a new one goes in. VideoLocked
8. 8Make every skipped step visibleUnder a push, the agent ran every step. Told twice to skip the review, it did, and the only trace was a commit message. Ask Claude Code to move the bug-fix workflow into an engine: it does not stop a developer who insists, but it keeps where the run stopped, why, and who sees it. VideoLocked
9. 9Certificate: govern your own repoThe course is a set of moves, not a set of files: each gives one kind of knowledge one home, a cap and a way to retire it, and each is checked by a run. Walk the eight in order, add one check over all of them, and take them to your own repo for the certificate. VideoLocked