Menu
Hypothesis-driven debugging method — reproduce, isolate, bisect, prove the cause before fixing. Apply whenever observed behavior differs from expected behavior. Independent of project, language, and domain.
Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
Turn a vague or ambiguous request into executable work — establish the facts, ask the right questions (and only those), present options with a recommendation, get explicit agreement before executing. Independent of project, language, and domain.
General execution discipline for any non-trivial task — decompose before acting, prove before concluding, decide next steps on facts, refuse shortcuts. Independent of project, language, and domain.
Based on SOC occupation classification
| name | scientific-debugging |
| description | Hypothesis-driven debugging method — reproduce, isolate, bisect, prove the cause before fixing. Apply whenever observed behavior differs from expected behavior. Independent of project, language, and domain. |
A bug is a gap between what you observe and what you expected. Debugging it means explaining that gap with a proven cause — not making it disappear. A symptom that vanishes without explanation is not a fixed bug, it's a hidden bug.
Write the gap in one factual sentence. "When I do X, I observe Y; I expected Z." If you can't fill in all three slots, you don't have a bug yet — you have an impression. Go get the three first.
Reproduce before explaining. Until you have triggered the bug yourself, you're working on a story, not a fact. A bug you can't reproduce can't be fixed — it gets trapped (logs, instrumentation) until it becomes reproducible.
Shrink the reproduction to the minimum. Remove everything that can be removed while the bug persists: data, steps, options, dependencies. Every element removed without making the bug disappear is exonerated; every element whose removal makes the bug disappear is a suspect. The minimal reproduction IS already half the diagnosis.
Read the error message literally. Word for word, including the file name, the line number, the error code, and the first error in the stack — not the last one, which is often just a consequence. The classic trap: reading what you expect the error to say instead of what it says. If a term in the message is unknown to you, look it up before forming any hypothesis.
Establish "since when". Did it ever work? If so, what changed between when it worked and now — code, dependency, environment, data, configuration? A regression is searched for in the delta, not in the whole system.
Form hypotheses BEFORE touching the code. Two or three plausible causes, each written with its testable prediction: "if it's A, then by doing T I should observe O". A hypothesis without a prediction is not a hypothesis, it's an intuition — unusable.
Test the cheapest hypothesis first, not the most seductive one. A hypothesis eliminated in 30 seconds beats a brilliant hypothesis that needs an hour of setup.
Change ONE variable at a time. If you change two things and the behavior changes, you don't know which one acted — you burned a trial and polluted your reasoning. Return to the reference state between trials.
Bisect the space of causes. Every test should cut the remaining space in half, whatever the dimension: in time (which version introduced the bug), in code (does the bug survive if I short-circuit this half of the path), in data (does the bug persist with half the input). Ten bisections cover a thousand suspects; ten random checks cover ten.
Look at the actual state, not the assumed state. Print the effective values at the point of interest: what goes in, what comes out, what is configured, which version is actually running. Half of all "incomprehensible" bugs are a gap between the system you think you're observing and the one that's running: wrong environment, non-invalidated cache, old deployed version, unsaved file. Check THAT point first when nothing makes sense.
When nothing makes sense, question your certainties, in order. List what you hold as given ("this function is called", "this file is the one being loaded", "this value is what I think it is") and verify them one by one, from the most fundamental to the finest. The bug always lives in a box labeled "obvious, no need to check".
Keep a log of your search. For any investigation beyond a few trials, keep a written trace: hypothesis → test → result → verdict. Without a trace, you will re-test the same hypothesis twice and call it perseverance.
Demand a complete explanation before fixing. You must be able to tell the causal chain end to end: "X triggers Y, which produces Z, which yields the observed symptom". If a link stays fuzzy ("and there, for some reason…"), you're not done searching.
A fix that works without you knowing why is a failure. If the symptom disappears after a change you can't tie to your causal chain, don't stop: either your theory is wrong, or the bug is masked and will come back. Remove the fix, verify the bug returns, put it back, verify it leaves — if that back-and-forth isn't clean, you fixed nothing.
Fix at the level of the contract, not the symptom. Forbidden, except with explicit and documented justification:
Each of these gestures turns a loud bug into a silent one — an aggravation, not a fix.
Hunt the bug's siblings. Does the cause you found exist elsewhere? The same pattern copied, the same API misused, the same false assumption on another path. A found bug is a free detector: run the code through the same filter before closing.
Add the test that would have caught the bug. The test must fail on the code from before and pass on the code from after — verify both. A test written after the fact that never failed proves nothing.
Give each lead a budget. If a hypothesis withstands several tests without being confirmed or eliminated, it is badly posed: reformulate it or change leads. Grinding on a single lead is the most common failure mode — it looks like work.
Escalation is a deliverable. If the investigation runs dry, deliver a structured report: exact symptom, minimal reproduction, hypotheses tested with their verdicts, what remains open and what information is missing. That report has value; "I didn't find it" has none. Whoever picks up after you must be able to start from your eliminations, not redo your search.
Never "fix" by accumulation. If you catch yourself stacking changes "to see what happens", stop everything, return to the clean reference state, and start over at section 1. A system riddled with exploratory modifications is harder to debug than the original one — you are manufacturing a second bug.
