| name | diagnose |
| description | Disciplined 6-phase bug diagnosis loop. Build feedback loop, reproduce, hypothesise, instrument, fix with regression test, cleanup with post-mortem. Use for all bugs — bypasses PRD entirely. Ends with advisor-gate completion gate. |
Diagnose
Core Principle
A disciplined loop, not guesswork. The feedback loop IS the skill — everything else is mechanical. Without a fast, deterministic pass/fail signal, no amount of code reading will help.
Verify the claim before any fix is delegated (triage gate). Reproduce the reported failure FIRST — run it, observe the actual symptom (Phases 1–2) — before routing a fix to a general-purpose. Never delegate a fix for a bug you have not reproduced: a fix for an unconfirmed claim is a guess, and guesses ship the wrong change. If you cannot reproduce it, do NOT proceed to a fix — stay in triage: sharpen the loop, gather evidence (logs, HAR, a failing input, environment access), or ask the user for the missing piece. "Can't reproduce" is a reason to investigate, not to guess-patch.
This skill replaces create-prd and implement for bugs. Bugs go through: interview (optional scoping) → diagnose → advisor-gate. No PRD needed.
When to Use
- Any reported bug or regression
- Performance degradation
- "It worked before, now it doesn't"
- User reports unexpected behavior
- Test failures that need root cause analysis
The 6 Phases
Phase 1 — Build a Feedback Loop
"This is the skill. Everything else is mechanical."
Construct a fast, deterministic, agent-runnable pass/fail signal. Without one, stop and ask for help.
10 ways to build a loop (priority order):
- Failing test at whatever seam reaches the bug
- HTTP script —
curl / HTTP client against dev server
- CLI invocation — fixture + diff stdout vs known-good snapshot
- Headless browser — Playwright/Puppeteer script
- Replay trace — captured network request, payload, or event log
- Throwaway harness — minimal subset of the system
- Fuzz/property test — 1000 random inputs
- Bisection harness — for
git bisect run
- Differential — old version vs new version comparison
- HITL script — human-in-the-loop bash script (last resort)
Iterate on the loop itself. Make it faster, sharper, more deterministic. A 2-second deterministic loop is a debugging superpower.
Non-deterministic bugs: Goal is higher reproduction rate (50% is debuggable; 1% is not). Loop 100x, parallelize, add stress.
When genuinely impossible to build a loop: Stop and say so. Ask for: (a) environment access, (b) captured artifact (HAR, log dump, core dump, screen recording), or (c) permission to add temporary production instrumentation.
Phase 2 — Reproduce
Run the loop. Confirm:
- Reproduces the user-described failure (not a different nearby one)
- Reproducible across runs
- Exact symptom captured (error message, wrong output, timing)
If the loop doesn't reproduce: refine the loop or go back to Phase 1.
Phase 3 — Hypothesise
Generate 3-5 ranked hypotheses before testing any. Single-hypothesis generation causes anchoring.
Each hypothesis must be falsifiable: "If is the cause, then will make the bug disappear / will make it worse."
Show the ranked list to the user before testing. User often has domain knowledge to re-rank instantly.
Parallel hypothesis testing: Launch the top 2-3 hypotheses as parallel exploration tasks when they probe different parts of the system:
# Good: parallel probes of independent hypotheses
# Both orchestrator and general-purpose can delegate to explore
task(description="Test hypothesis A: auth middleware", prompt="Check if the auth middleware strips the X-Token header when...", subagent_type="explore")
task(description="Test hypothesis B: race condition in cache", prompt="Check if the cache invalidation runs before the response...", subagent_type="explore")
Only parallelize when hypotheses are independent. If Hypothesis B depends on Hypothesis A being wrong, test A first.
Phase 4 — Instrument
Each probe maps to a specific prediction from a hypothesis. Change one variable at a time.
Tool preference (in order):
- Debugger / REPL
- Targeted logs at hypothesis-distinguishing boundaries
- Never "log everything and grep"
Tag every debug log with unique prefix [DEBUG-xxxx] — cleanup is a single grep.
Performance branch: For performance regressions, logs are usually wrong. Establish baseline measurement first.
Phase 5 — Fix + Regression Test
Write regression test before the fix — at a correct seam (exercises the real bug pattern as it occurs at the call site).
If no correct seam exists, that itself is the finding — flag for architecture improvement and note in post-mortem.
Parallel execution: Write the regression test AND the fix simultaneously:
# Write the failing test first, then apply the fix in the same task
# The test and fix touch the same files, so they must be in the same task
# But you CAN parallelize: fix implementation + feedback loop verification
For orchestrator: Delegate the fix to a general-purpose agent and verification to explore:
task(description="Write regression test + apply fix", prompt="...", subagent_type="general-purpose")
task(description="Verify feedback loop passes after fix", prompt="...", subagent_type="explore") # runs after fix task
For general-purpose: Implement the fix and regression test yourself. You can delegate codebase verification to explore via task(subagent_type="explore", ...).
Sequence:
- Minimise reproduction → write failing test → watch it fail
- Apply fix → watch test pass
- Re-run original feedback loop → confirm fix
Phase 6 — Cleanup + Post-Mortem
Checklist:
Post-mortem question: What would have prevented this bug?
If answer involves architectural change (no good test seam, tangled callers, hidden coupling), note it for future architecture improvement work.
Abbreviated Mode
For trivial bugs where the cause is obvious (≤1 file, obvious fix):
Skip directly to: Phase 1 (quick loop) → Phase 2 (reproduce) → Phase 5 (regression test + fix) → Phase 6 (cleanup)
Skip Phase 3 (hypothesise) and Phase 4 (instrument) — there's only one plausible cause and it's already identified.
Completion Gate
After Phase 6, invoke advisor-gate with:
- Original bug report
- Root cause (which hypothesis was correct)
- Fix summary
- Regression test location
- Post-mortem finding
This is non-negotiable. Diagnose always ends with advisor-gate.
What NOT to Do
- Do NOT skip building a feedback loop — guessing at code is not diagnosis
- Do NOT test hypotheses one at a time as they're generated — generate all 3-5 first, rank, then test
- Do NOT write the fix before the regression test
- Do NOT leave
[DEBUG-xxxx] instrumentation in the code
- Do NOT invoke
create-prd or implement for bugs — this skill owns the entire bug path
- Do NOT skip the
advisor-gate completion gate
