メニュー
Diagnose software failures through evidence-based root-cause analysis. Use for bugs, failing tests, runtime errors, data anomalies, stuck jobs, environment failures, and intermittent or expensive-to-reproduce incidents.
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
SOC 職業分類に基づく
Bootstrap software repositories from live evidence. Use for initial setup, inherited-codebase rehabilitation, repository mapping, conventions, documentation architecture, safe restructuring, or agent-ready tooling.
Solve ambiguous, high-stakes, or multi-variable questions with a labelled problem tree. Use for open-ended life, business, product, strategy, and root-cause reasoning; route concrete feature briefs to $grill-me and specific failures to $debugging.
Create, edit, and run persistent Codex workflow graphs with optional GitHub delivery.
Turn a rough software feature request into one codebase-grounded implementation brief. Use for feature pitches, vague product requirements, solution exploration, or explicit “grill me” requests before coding; use $brainstorming for non-implementation questions.
Implement an agreed story, plan, or failing-test target and prove the resulting behavior. Use when an accepted contract authorizes production changes; route diagnosis-only work to $debugging and test design to $tdd.
Land finished work in a Git repository through Gate, Record, Commit, and Publish. Use when the user asks to close completed work, commit and push it, or open a pull request; use repository review workflows for reviewing an existing PR.
| name | debugging |
| description | Diagnose software failures through evidence-based root-cause analysis. Use for bugs, failing tests, runtime errors, data anomalies, stuck jobs, environment failures, and intermittent or expensive-to-reproduce incidents. |
Read the repository's debugging instructions and known-failure records when present; repository- specific procedures own commands and evidence sources.
Verify the symptom, find the first divergence from expected behavior, identify the violated invariant, compare competing hypotheses, seek discriminating evidence, and derive the causal chain. When reproduction is costly, trace backward through callers, contracts, state transitions, async boundaries, concurrency, retries, caches, configuration, and partial failures.
Keep facts, inferences, and unknowns distinct. Separate root cause from trigger, contributing factor, and downstream symptom. Try to falsify the leading hypothesis and explain the strongest rejected alternative. Runtime evidence supports direct conclusions; static-only reasoning receives lower confidence and an explicit confirmation path.
When available evidence cannot distinguish the surviving hypotheses, return inconclusive with the
ranked candidates, shared evidence, and smallest discriminating observation still needed. An
inconclusive result names no root cause.
This skill's deliverable is diagnosis. Implement or recommend a fix only when the user separately authorizes that scope.
Report in at most 180 words unless detail is requested:
Diagnosis is complete when one causal chain explains the observed symptom, survives the strongest available falsification attempt, and every material uncertainty is named. The inconclusive branch is complete when every surviving hypothesis and the exact evidence needed to distinguish them are named without promoting one to a diagnosis.
