Thank you for creating this! Having a way to force AI to do TDD has made me much more open to letting it write code with less supervision (as an actual "agent", as opposed to me just using it during a spike or for troubleshooting).
My team and I have been using it for the last month or so, and have made a few tweaks that have gotten some mileage on them now, and if you're open we can merge them into your upstream here.
One thing we found was that despite the clear guidance to only write the minimal amount of code needed to make the test pass, Claude is still a bit overeager sometimes.
We added this to the green.md file right after "Address the specific failure message", and we've found it effective:
### Post-Green Verification
Once the test passes, run the coverage tool scoped to **only the files you edited** and check for uncovered lines:
- Any uncovered lines in files you edited are over-implementation — **delete them**
- Do not scope to the full test suite; focus only on what changed
Thank you for creating this! Having a way to force AI to do TDD has made me much more open to letting it write code with less supervision (as an actual "agent", as opposed to me just using it during a spike or for troubleshooting).
My team and I have been using it for the last month or so, and have made a few tweaks that have gotten some mileage on them now, and if you're open we can merge them into your upstream here.
One thing we found was that despite the clear guidance to only write the minimal amount of code needed to make the test pass, Claude is still a bit overeager sometimes.
We added this to the
green.mdfile right after "Address the specific failure message", and we've found it effective: