Debugging

Engineering quality focused skills for AI coding agents that keep humans in the loop.

Install

npx -y skills add kreek/consult --skill debugging

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

One thing to look at

1 stars1 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

Copied from the file, not written here

Use to debug failures, reproduce symptoms, isolate causes, inspect evidence, and fix bugs.

3.8 KB, as published. Nobody here has run it

NO FIX WITHOUT ROOT-CAUSE EVIDENCE.

Investigating defects, flakes, timing-sensitive bugs, regressions, production incidents, unclear failures, or "stuck" debugging sessions.

Reproduce before fixing. Change one variable per experiment.
Reduce the failing case until only the bug remains. Localize by boundary: data, service, integration, application, infrastructure.
Debugging should improve the human's failure model with evidence.
Fix causes, not symptoms.
For incidents, produce blameless learning with owned follow-up actions, never "human error" as a root cause.

Capture the exact symptom: command, input, output, stack trace, timing, environment. Reproduce reliably or record why reproduction isn't yet possible.
After the third experiment, keep a short debug log. Form one hypothesis at a time, predict what else must be true, run the smallest experiment that confirms or refutes it.
Before editing, state the current failure model: likely cause, evidence for it, and the observation that would disprove it.
Fix the identified cause. Add a regression test or operational guard before declaring fixed.

The bug reproduces on pre-fix code or non-reproducibility is documented.
The root cause is named in one sentence, explains all observed symptoms, and identifies evidence that ruled out main alternatives.
The fix is one atomic change aimed at that cause, with a regression test or equivalent guard that fails before and passes after.
Timing-sensitive bugs were verified with non-invasive observation or replay.
Incident follow-ups have owners and deadlines.
Unproven root-cause or fix claims are reported as unproven.

Trigger	Do this instead	False alarm
"Probably X, let me try fixing"	Gather evidence that confirms X before editing.	The user asked for a speculative explanation, not a fix.
"No time to reproduce"	Create the smallest reproduction or state why reproduction is blocked.	Production-only incident where logs/traces are the available reproduction.
"One more guess and it'll work"	Stop editing. Collect a new observation that changes the model.	A syntax or wiring typo found directly in the failing output.
"The agent found a fix but I can't explain the bug"	Stop and rebuild the failure model before editing further.	Disposable local experiment with no completion claim.
"Fixed it locally, ship it"	Name the root cause and add or run the regression guard.	Local run is the requested diagnostic, not a completion claim.
"Flake - just retry"	Treat the flake as a bug and identify whether test, code, or environment failed.	Infrastructure outage already confirmed outside the code under review.
"Probably a race condition"	Show interleaving, shared state, or timing evidence before changing concurrency code.	The race is already demonstrated by a failing sanitizer or trace.

Use proof for fix-claim evidence and the regression test shape.
Use git-workflow for git bisect, reflog recovery, or conflict-heavy debugging.
Use observability when the right evidence must come from logs, metrics, traces, or incident timelines.
Use error-handling when the root cause is an error boundary, retry, timeout, or recovery contract.