Systematic Debugging

Overview

When LLMs hit failures, they tend to guess at fixes — changing code, re-running, changing more code. This thrashing wastes context and often makes things worse. This skill enforces a four-phase protocol that requires understanding before action. Systematic debugging resolves issues in 15-30 minutes; guess-and-fix typically takes 2-3 hours with a 40% first-attempt success rate vs 95% systematic.

Iron Law

Find root cause before attempting fixes. Symptom fixes are failure. No fixes without root cause investigation first.

Constraints

• Do not change code until you have a root cause hypothesis
• Do not apply more than one fix at a time
• Do not skip the reproduction step — a bug you can't reproduce is a bug you can't verify as fixed
• Do not guess. Investigate.
• After 3+ failed fix attempts, question the architecture — stop patching

The Four Phases

Phase 1: Reproduce

Goal: See the failure with your own eyes (tool output).

1. Run the failing test or trigger the error condition
2. Read error messages thoroughly — don't skim. Stack traces contain line numbers, file paths, and often the solution.
3. Identify the minimal reproduction — strip away everything unrelated
4. Determine if the issue is consistent or intermittent
5. Gate: paste the reproduction output. If you can't reproduce it, gather more data — don't speculate.

Phase 2: Investigate & Pattern Analysis

Goal: Gather facts and find working reference points.

Investigate (use as many techniques as needed):

• Read the error: Parse the stack trace. What file, line, and function? What's the actual error type?
• Trace the data flow: Follow the input from entry point to failure point. Where does the actual value diverge from the expected value?
• Check recent changes: What changed since this last worked? (git diff, git log)
• Add observation points: Temporary logging or print statements at key points to see actual values
• Multi-component systems: Add diagnostics at each component boundary — log data entering and exiting each layer to find which layer fails
• Bisect: If the failure surface is large, narrow it down by testing midpoints

Pattern Analysis:

• Find working examples: Locate similar functioning code in the same codebase
• Compare against references: Read reference implementations completely — don't skim
• Identify differences: List every distinction between working and broken code, no matter how minor
• Understand dependencies: What components, settings, and assumptions are required?

Gate: state what you know and don't know. List the facts, not guesses.

Phase 3: Root Cause Hypothesis

Goal: Identify the single underlying cause using the scientific method.

1. Form a single hypothesis: "I think X is the root cause because Y"
2. Predict what you would see if the hypothesis is correct (a test you haven't run yet)
3. Run that prediction test — make the smallest possible change, one variable at a time
4. If the prediction fails, your hypothesis is wrong — return to Phase 2 with new information
5. Do not compound fixes — test one hypothesis at a time
6. Gate: state the root cause in one sentence. If you can't, you don't have it yet.

Phase 4: Fix

Goal: Make the smallest change that addresses the root cause.

1. Write or modify a test that captures the bug (it should fail now)
2. Apply the fix — one change, targeting the root cause
3. Run the test — it should pass
4. Run the full suite — no regressions
5. Gate: paste the test output showing the fix works and nothing else broke

If the fix doesn't work: Stop and reassess. After fewer than 3 attempts, return to Phase 1. After 3+ failures, question the architecture itself — when each fix reveals new problems elsewhere, the bug is architectural, not local. Discuss fundamentals with the human before attempting more fixes.

Red Flags Requiring Process Restart

Stop immediately and return to Phase 1 if you notice:

• Planning "quick fix for now, investigate later"
• Attempting changes without understanding why they'd work
• Adding multiple modifications simultaneously
• Skipping testing in favor of "just checking manually"
• Fixing without investigation
• Not understanding something but proceeding anyway
• Adapting patterns differently than reference implementations
• Proposing solutions before tracing data flow
• Attempting a fourth or subsequent fix on the same issue
• Each fix revealing new problems in different locations

Rationalization Prevention

Excuse	Reality
"I think I know what's wrong, let me just try this"	That's guessing. Investigate first — it takes less time than three wrong guesses.
"The fix is obvious from the error message"	Then it'll be fast to verify your hypothesis before coding. Do it.
"I'll just try a few things and see what sticks"	Each attempt burns context and may introduce new bugs. One investigated fix beats five guesses.
"This is probably a race condition / timing issue"	"Probably" isn't a root cause. Add observation points and prove it.
"Let me revert everything and start over"	You'll hit the same bug again. Understand it first, then decide whether to revert.
"It's a simple issue, I don't need the full process"	Root causes exist in simple bugs too. The process handles simple issues quickly — just do it.
"We're under time pressure, just fix it fast"	Systematic debugging actually beats guess-and-check even under time pressure.
"Try fixing first, investigate if it doesn't work"	How you start establishes the pattern. Start correctly.
"Multiple simultaneous fixes save time"	You can't isolate what worked. You'll create new bugs. One fix at a time.
"I can see the problem, the cause is obvious"	Symptoms differ from root causes. What you see is the symptom. Investigate the cause.

Success Indicators

Phase	You're Doing It Right When...
Reproduce	You have consistent reproduction with exact error output
Investigate	You can list facts (not guesses) about what's happening and have working reference code
Root Cause	You have a confirmed or revised hypothesis with prediction test results
Fix	Issue is resolved with a passing test and no regressions

When to Use This Skill

• A test that was passing now fails
• An error occurs during implementation
• Unexpected behavior is observed
• A bug report comes in
• A fix you applied didn't work (re-enter at Phase 2)
• You're under time pressure (especially then)

Output

Root cause analysis with evidence: reproduction output, investigation findings, root cause statement, fix applied, and verification output showing the fix works without regressions.