You tell your AI agent to use pnpm, not npm. Tomorrow you tell it again. By Friday you've typed the same correction for the third time, and your AGENTS.md still says nothing about pnpm.
Your AI config is a product. It has users (you and your team), it gets bug reports (every correction you type), and it should ship fixes. Every bit of friction should end as a small change to the config, so the same mistake never costs you twice.
Anything that changes how your agent behaves: AGENTS.md / CLAUDE.md, skills, hooks, plugins and MCP servers. See what each layer does on Do you customize your coding agent?
Most people write an AGENTS.md, grab a skill from their co-worker, or create a hook once on day one and never open it again... but they still use it every single day.
The codebase moves on. Team habits change. Your agent config goes stale, and you start patching the gap by hand in every chat.
"They're alive - they evolve with time and usage. They're yours - they're shaped by your workflow and your frictions." - Josh Berman, Closing the Loop on the Software Factory
Use the same idea as a retrospective. A retro only works when it ends with an action point. For an AI agent, the action point is a config change.
Without a retro loop - You correct the agent in chat every session. The config never changes, so tomorrow's session makes the same mistake.
❌ Figure: Bad example - Corrections live in chat and disappear when the session ends
With a manual retro - After a frustrating session, you run a retro skill like Matt Pocock's retro skill. It proposes a fix and you add it. This works, but only when you notice the friction and remember to run it.
🙂 Figure: OK example - Fixes land, but only when you remember
With a nightly retro - A scheduled "night shift" job reads the day's agent sessions. It logs every correction, rejected tool call and revert to a ledger. When a pattern repeats across sessions, it writes a proposal: update a skill, add a hook, or add a one-liner to AGENTS.md. Each morning you implement or skip each proposal.
✅ Figure: Good example - Friction is caught automatically and a human approves every fix
Pick the lightest fix that stops the friction:
AGENTS.md, e.g. "Use pnpm, never npm"Warning: Never auto-apply proposals. Pattern matching can lock in a bad habit as easily as a good one. Review every config change the same way you review code.
Tip: Commit project config to the repo and review changes in PRs, so every fix reaches the whole team. Keep personal preferences in CLAUDE.local.md so they don't leak into the shared file.
If the loop is automated, review the proposals every morning. It takes 2 minutes.
If it's manual, run a retro after any session where you corrected the agent twice. Then run /insights monthly for the bigger picture.
However this should be suited how often you are using AI. For those using it constantly every day, improving your config every day can have a compounding impact on your workflows.