Self-annealing
Build-along · ~12 min at your own keyboard
Hire employee B: when something breaks, they fix it AND update the manual. Add a standing rule to your workspace — on any error: diagnose, fix, update the instructions, and log the change. Then prove the fix: re-run the input that broke it, plus one similar case. A changed instruction is a hope until a test passes. Keep the old version so you can roll back if the fix breaks something else.
The long version: Self-annealing workflows ↗
Do this
0/4Checkpoints
- ▸You should now see the annealing rule inside CLAUDE.md.
- ▸You should now see a failure trigger a fix, a doc update, and a changelog line.
- ▸You should now see the original failing input and one similar case both pass on a re-run.
The three ways this goes wrong
- Claude fixes the error but does not update the docs. — Make the rule explicit: 'after any fix, edit the instructions and append a line to CHANGELOG.md.'
- The changelog fills with noise. — One line per real fix: date, what broke, what changed.
- The fix passes the original case but breaks something else. — Roll back to the saved instructions, then fix again with both cases as tests.
Share what happened
Paste your annealing rule, the changelog line from a real break, the doc line that changed, and the result of re-running the broken input plus one similar case.
Strip API keys, passwords, and tokens before you paste. Swap them for [KEY].
🔒 Don't paste confidential, personal-health, or payment info.