Problem
AI gets stuck because it's building on unverified assumptions about the code.
- Assumes functions return X when they return Y
- Misinterprets errors based on a wrong mental model
- Each assumption becomes a foundation for the next wrong step
This is a human problem too - when you can't run code frequently but keep writing it, you make assumption chains. AI does the same.
What Goes Wrong
AI gets stuck spinning:
- Checks step A, step C, step V
- Misses the wrong assumption at step B
- Every "fix" builds on that wrong foundation
Solution
- When AI gets stuck, stop it and tell it to validate each step incrementally
- Use TDD to create automatic micro-feedback loops that catch drift early. Try Predictive TDD - AI predicts test outcomes, gets surprised when wrong (like humans do), immediately corrects its mental model
The key: frequent reality checks prevent assumption chains. Code's self-verifiable nature makes this possible in ways that wouldn't work for general facts.
Example
AppleScript debugging:
AI kept adjusting patterns, tweaking parsing, trimming whitespace - all based on one hidden assumption: AppleScript's log returns the session IDs.
Wrong. log prints descriptions, doesn't return values. Needed return instead. Because it never validated that first step, every downstream fix was wasted effort.
Asked "check each command step by step" - it isolated the AppleScript call, validated output, saw the real problem. Fix was trivial once the base assumption was checked.
This happens with humans too: Early in my career, paired with someone who wrote PL/SQL for three hours without running it once. So many unvalidated assumptions. When we finally ran it, it didn't even compile.
