daBongo LMS AI Training Courses

Lesson Content

The behavior-first debugging protocol.

Step 1 – Define the failure precisely:

• What is the observed behavior? (exact error, wrong output, wrong state)
• What is the expected behavior?
• Under what input or conditions does the failure occur?
• Is the failure consistent or intermittent?

Step 2 – Share relevant context:

• The code path that handles the failing case (file + function or component)
• The error trace or logs (full stack trace where available)
• Any recent changes (what changed before the failure appeared?)

Step 3 – Ask for hypotheses, not fixes:

• "What are the most likely causes of this behavior given this code and this error?"
• Review hypotheses against your knowledge of the system
• Test the most likely hypothesis before asking for a fix

This sequence produces diagnosis. Skipping to "fix this" produces guesses.

Log and error trace analysis.

Claude Code reads logs and stack traces effectively. When sharing:

• Include the full error message and stack trace – not just the error type
• Include the relevant lines of code, not just the trace
• For async errors, include the call site and the error location

Ask Claude to interpret a stack trace before asking for a fix: "What does this stack trace tell us about where the failure occurred and what caused it?" This produces understanding, not just a patch.

Hypothesis testing workflow.

For each hypothesis Claude generates:

1. Evaluate: does this hypothesis match the symptoms you observed?
2. Design a test: how would you verify this hypothesis?
3. Test: make the minimal change to test the hypothesis
4. Evaluate: did the behavior change as predicted?

This is standard scientific debugging. Claude accelerates step 1 (generating hypotheses) – the developer still tests and evaluates.

When debugging requires human expert knowledge.

Claude is well-suited for: common bug patterns, syntax errors, async/await misuse, off-by-one errors, type errors, common library misuse.

Claude is less reliable for: performance profiling in specific hardware or runtime contexts, database query optimization for specific data distributions, distributed systems race conditions, domain-specific edge cases with limited training data coverage.

For performance and distributed systems bugs, Claude's hypotheses are a starting point – not a conclusion.