菜单
Use when diagnosing bugs, failing tests, failed pipelines, bad data, or unexpected runtime behavior. Do not use for routine code review or simple formatting tasks.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Use when implementing or reviewing AV3 candidate-RAG work, including retrieval, Com_RAG candidate answers, context snapshots, candidate prompt rendering, candidate execution, or Sem_RAG vs Com_RAG comparison. Do not use for AV2-only judge pipeline work unless adapting judges to AV3 candidate answers.
Use when implementing or reviewing the AV2 LLM-as-a-Judge execution pipeline. Do not use for generic Python setup, standalone SQL analysis, or rubric design without execution changes.
Use when creating, documenting, or validating PostgreSQL backup and restore artifacts for AV2. Do not use for judge execution, rubric design, or Python project setup.
Use when importing datasets, AV1 answers, prompts, rubrics, or judge outputs into PostgreSQL and validating referential integrity. Do not use for judge prompt design or pure Python project setup.
Use when designing or reviewing LLM-as-a-Judge rubrics and scoring prompts for AV2. Do not use for database imports, backup/restore, or generic Python setup.
Use when creating, reviewing, or standardizing Python project infrastructure. Do not use for domain-specific implementation, business logic, or application feature design.
| name | defect-root-cause-analysis |
| description | Use when diagnosing bugs, failing tests, failed pipelines, bad data, or unexpected runtime behavior. Do not use for routine code review or simple formatting tasks. |
Diagnose defects systematically using evidence, reproduction, data-flow tracing, and targeted validation.
Use this skill when the task involves:
Do not use this skill for:
Gather:
Start with inspection:
git status --short
git diff --stat
For tests:
.venv/bin/python -m pytest path/to/test.py -q
.venv/bin/python -m pytest
For logs:
grep -R "ERROR\|FAILED\|Traceback" path/to/logs
Adjust commands to the project.
Run the narrowest command that reproduces the failure.
Do not fix before confirming the failure unless reproduction is impossible.
Classify as:
Map:
Identify:
Use evidence:
Apply the narrowest change that fixes the proven root cause.
Do not refactor stable code unless necessary.
Run:
## Context & Goal
## Reproduction
## Failure classification
## Data flow
## Code path
## Root cause
## Fix
## Validation performed
## Risks and follow-ups