메뉴
Manage eval-driven development workflow with define, check, report, and list commands. Use when user says "define eval", "check evals", "eval report", or "list evals". Tracks pass@k metrics for capability and regression evals.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
SOC 직업 분류 기준
| name | eval |
| description | Manage eval-driven development workflow with define, check, report, and list commands. Use when user says "define eval", "check evals", "eval report", or "list evals". Tracks pass@k metrics for capability and regression evals. |
Manage eval-driven development workflow.
/eval [define|check|report|list] [feature-name]
/eval define feature-name
Create a new eval definition:
.claude/evals/feature-name.md with template:## EVAL: feature-name
Created: $(date)
### Capability Evals
- [ ] [Description of capability 1]
- [ ] [Description of capability 2]
### Regression Evals
- [ ] [Existing behavior 1 still works]
- [ ] [Existing behavior 2 still works]
### Success Criteria
- pass@3 > 90% for capability evals
- pass^3 = 100% for regression evals
/eval check feature-name
Run evals for a feature:
.claude/evals/feature-name.md.claude/evals/feature-name.logEVAL CHECK: feature-name
========================
Capability: X/Y passing
Regression: X/Y passing
Status: IN PROGRESS / READY
/eval report feature-name
Generate comprehensive eval report:
EVAL REPORT: feature-name
=========================
Generated: $(date)
CAPABILITY EVALS
----------------
[eval-1]: PASS (pass@1)
[eval-2]: PASS (pass@2) - required retry
[eval-3]: FAIL - see notes
REGRESSION EVALS
----------------
[test-1]: PASS
[test-2]: PASS
[test-3]: PASS
METRICS
-------
Capability pass@1: 67%
Capability pass@3: 100%
Regression pass^3: 100%
NOTES
-----
[Any issues, edge cases, or observations]
RECOMMENDATION
--------------
[SHIP / NEEDS WORK / BLOCKED]
/eval list
Show all eval definitions:
EVAL DEFINITIONS
================
feature-auth [3/5 passing] IN PROGRESS
feature-search [5/5 passing] READY
feature-export [0/4 passing] NOT STARTED
$ARGUMENTS:
define <name> - Create new eval definitioncheck <name> - Run and check evalsreport <name> - Generate full reportlist - Show all evalsclean - Remove old eval logs (keeps last 10 runs)Batch dispatch — read queued tasks from ~/.claude/batch/inbox.md, fix each task's spec contract, fan them out to background subagents per case, and leave the user free until review time. Also surfaces unreported client sessions (~/.claude/batch/unreported.md). Use when the user says "batch", "morning batch", "朝バッチ", "タスク投入", "/batch", "/morning-batch", or at the start of a work day with queued tasks.
Standardized "investigate → written report" pipeline for SRE client work. Fixes the deliverable contract up front (format + done-criteria), delegates repo/infra scanning to subagents, verifies every claim, and outputs an issue-ready Markdown report. Use when the user says "調査して", "調べてレポート", "investigation report", "報告書にまとめて", or hands over an investigation-shaped task.
Turn a finished case (case-reflect / sreaas report output) into a generalized, confidentiality-scrubbed company tech-blog draft. Use when the user says "会社ブログ", "テックブログ", "案件を記事化", "company blog", or wants to asset-ify case outcomes for public publishing. Realizes the company-blog stage of learn_cycle.md.
Generate a personalized learning guide (text + hands-on script) for a technology/topic, tuned to the user's level and case context — like an AI-personalized intro book. Use when the user says "学習ガイド", "入門ガイド", "ハンズオンを作って", "learn guide", "この技術を学びたい", or wants a guided, output-oriented way to learn something. Powers the learning half of learn_cycle.md.
From a Claude Code /insights report, generate a Japanese-translated HTML report AND reflect its actionable recommendations into the config (~/.claude/), confirming EACH item one-by-one (apply or skip). Skipped items are left untouched. Use when the user says "insights apply", "insightを設定に反映", "insightを日本語HTML", "apply insights to config".
Systematically investigate and fix a GitHub Issue with root cause analysis. Use when user says "fix issue", "fix