Theo's tstack: the agent skills he uses to review and merge code

Try this

1. Write skill descriptions as triggers (when to use it), not as summaries. 03:56
2. Tell review agents to make the smallest possible change, to stop scope creep. 05:28
3. Cap review loops: three rounds by default, five for security or login changes. 05:46
4. Have the agent note in the PR description how many review rounds it ran and with which model. 07:18
5. Route model logins and quotas through one proxy (he uses CLIProxyAPI) so you can swap models inside command-line tools. 08:08
6. Have different models review the same change in parallel; they have different blind spots. 09:47
7. Treat the agent sidebar as an inbox and auto-archive threads once the work is merged. 12:35
8. Require concrete proof (command output, tests, screenshots) before any autonomous merge. 16:48
9. Add a 'when to ask me' section listing the only cases in which the agent may interrupt you. 18:08
10. When taking over someone else's PR, keep their commits and co-author credit. 24:50

Timestamps point to the moment in the source video.