Getting Claude to check its own coding work
Claude can get stuck in a loop where it changes code, but the change does not actually work. A person then has to suggest a specific check, such as running code to see whether one value becomes larger after a certain action. After setting up and running that check, Claude realizes that its own change failed.
The main question is how to make Claude think of a useful work check and run it automatically after making a change. One possible direction is to put that rule in the , so Claude treats testing as part of the job from the start.
Key points
- Claude may change code without confirming that the change works.
- The person has to suggest a concrete test before Claude checks the result.
- Once Claude runs the test, it can discover that its own fix failed.
- Adding a testing rule to the may help.
- For , should be part of the task, not an afterthought.