diagnose

Iron law: no fix without a reproduction and a confirmed cause. Don't shotgun changes hoping one sticks.

Work the loop:

1. Reproduce. Get a reliable, ideally deterministic repro before touching anything. If you can't reproduce it, that is the first problem to solve (control inputs, seeds, environment, timing).
2. Minimise. Shrink to the smallest failing case — strip away unrelated code, data, and config until only the essentials that still fail remain. Each thing removed is a variable eliminated.
3. Hypothesise. State a single, falsifiable hypothesis about the cause ("the cache returns a stale value because the key omits the tenant id"). One at a time.
4. Instrument. Add a log, assertion, breakpoint, or probe that will confirm or deny the hypothesis. Let the evidence decide — never guess-patch before the cause is proven. If the evidence refutes the hypothesis, form the next one and repeat.
5. Fix the root cause, not the symptom. If you're tempted to patch a symptom, name the root cause you're choosing not to fix and why.
6. Regression test. Add a test that fails before the fix and passes after, so the bug can't return. (Hand the failing-test-first part to the tdd skill.)

For flaky/intermittent bugs, the repro step is the work: pin down the nondeterminism (ordering, concurrency, time, external state) before hypothesising. Report what you ruled out, not just the fix.