| name | diagnosing-bugs |
| description | Disciplined diagnosis loop for hard bugs and performance regressions. Use when something is broken, throwing, failing, flaky, or slow, or when the user says debug, diagnose, or "why is this happening". |
Diagnosing bugs
A discipline for bugs that did not fall to the first read. The whole skill is Phase 1; everything after it is mechanical.
Skip a phase only when you can say why.
Handling what the loop captures
Diagnosis means running commands and capturing their output. Two things must never end up in what you show, save, log, or paste into an issue or PR:
- Secrets. Write
<REDACTED> in their place. Build loops against environment variables so the credential stays in the environment rather than in the command you display. Captured artifacts carry auth headers — quote only the lines carrying signal.
- Regulated and user-authored content. Never user-entered text, transcribed or dictated text, model-generated content, or person-identifying fields. This is default-deny, not a denylist: quote opaque IDs, enums, counts, timings and error codes, and drop everything else. Prose under an innocuous key still leaks — a denylist of field names cannot catch it.
Two consequences worth stating outright, because they are where this usually goes wrong:
- Build the repro on synthetic data. A fixture that reproduces the bug is a better artifact than a real record anyway — it is redistributable, it can go in the regression test, and it cannot leak.
- Temporary production instrumentation is the last resort, and it logs identifiers only. Adding a log line that dumps a free-form value into a telemetry sink is the exact failure a default-deny logging rule exists to prevent. If the diagnosis genuinely requires seeing a real value, that is a conversation with the user, not a patch.
If redaction leaves you without enough signal to diagnose, say so and ask.
Phase 1 — build a feedback loop
This is the skill. If you have a tight pass/fail signal that goes red on this bug, you will find the cause — bisection, hypothesis testing and instrumentation all just consume it. Without one, no amount of reading code will save you.
Spend disproportionate effort here. Be aggressive, be creative, refuse to give up.
Ways to build one, roughly in order
- Failing test at whatever seam reaches the bug.
- HTTP script against a running dev server.
- CLI invocation with a fixture input, diffed against known-good output.
- Headless browser script driving the UI, asserting on DOM, console or network.
- Replay a captured payload — save a request or event to disk (redacted, synthetic where possible) and replay it through the code path in isolation.
- Throwaway harness — a minimal subset of the system, dependencies mocked, that hits the bug path in one function call.
- Property or fuzz loop — for "sometimes wrong output", run many random inputs and look for the failure mode.
- Bisection harness — if it appeared between two known states, automate "boot at state X, check, repeat" so
git bisect run can drive it.
- Differential loop — same input through two versions or two configs, diff the outputs.
- Scripted manual loop — last resort, when a human must click. Give them an exact numbered script and a precise description of what to capture, so the result still comes back as structured evidence rather than an impression.
Tighten it
Once you have a loop, treat it as a product and make it tight:
- Faster — cache setup, skip unrelated init, narrow the scope.
- Sharper — assert on the specific symptom, not "did not crash".
- More deterministic — pin time, seed randomness, isolate the filesystem, freeze the network.
A 30-second flaky loop is barely better than nothing. A 2-second deterministic one is a superpower.
Non-deterministic bugs
The goal is not a clean repro, it is a higher reproduction rate. Loop the trigger, parallelise, add stress, narrow timing windows, inject sleeps. A 50%-flake bug is debuggable; a 1% one is not. Keep raising the rate until it is.
Phase 1 is done when
You can name one command you have already run at least once — show the invocation and its output, redacted — that is:
If you catch yourself reading code to build a theory before that command exists, stop. Jumping to a hypothesis is the exact failure this skill prevents. No red-capable command, no Phase 2.
When you genuinely cannot build one
Stop and say so. List what you tried, and ask for one of: access to an environment that reproduces it, a redacted artifact (log dump, HAR, core dump, screen recording with timestamps), or a decision about instrumentation. Proceed to hypothesising without a loop and you are guessing with extra steps.
Phase 2 — reproduce, then minimise
Run the loop. Watch it go red. Confirm:
Then minimise: shrink to the smallest scenario that still goes red. Cut inputs, callers, config, data and steps one at a time, re-running after each cut.
This is not tidiness. A minimal repro shrinks the hypothesis space in Phase 3 and becomes the regression test in Phase 5.
Done when every remaining element is load-bearing — removing any one turns the loop green.
Phase 3 — hypothesise
Generate 3–5 ranked hypotheses before testing any of them. Generating one at a time anchors you on the first plausible idea.
Each must be falsifiable — state the prediction:
If X is the cause, then changing Y makes the bug disappear / changing Z makes it worse.
If you cannot state the prediction, it is a vibe. Sharpen it or drop it.
Show the ranked list before testing. The user often re-ranks it instantly ("we deployed a change to #3 yesterday") or has already ruled one out. Cheap checkpoint, large payoff. Do not block on it — proceed with your ranking if they are away.
Phase 4 — instrument
Every probe maps to a specific prediction from Phase 3. Change one variable at a time.
- Debugger or REPL where the environment supports it — one breakpoint beats ten logs.
- Targeted logs at the boundaries that distinguish the hypotheses.
- Never "log everything and grep".
Tag every debug log with a unique prefix — [DEBUG-a4f2]. Cleanup becomes one grep. Untagged logs survive forever; tagged logs die.
Performance regressions branch here. Logs are usually the wrong tool. Establish a baseline measurement — timing harness, profiler, query plan — then bisect against it. Measure first, fix second.
Phase 5 — fix, with a regression test
Write the regression test before the fix, if there is a correct seam for it.
A correct seam exercises the real bug pattern as it occurs at the call site. If the only reachable seam is too shallow — a single-caller test when the bug needs several callers, a unit test that cannot reproduce the triggering chain — a test there gives false confidence.
If no correct seam exists, that is itself the finding. Note it: the architecture is preventing the bug from being locked down.
Where a seam exists:
- Turn the minimised repro into a failing test there.
- Watch it fail.
- Apply the fix.
- Watch it pass.
- Re-run the Phase 1 loop against the original, un-minimised scenario.
Phase 6 — clean up and close out
Before declaring done:
Then ask what would have prevented it. If the answer is a missing test seam, tangled callers, or hidden coupling, that is a cleanup with a scope of its own — file it, or hand it to /refactorpass on the follow-up PR with the specifics. Make that call after the fix is in; you know more now than when you started.
Adapted from the diagnosing-bugs skill in mattpocock/skills (MIT), with default-deny telemetry rules replacing its generic redaction guidance — see NOTICE.
