菜单
Formal evaluation framework for Claude Code sessions implementing eval-driven development (EDD) principles
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 after a complex task, failure, or when reviewing what was learned. Teaches how to write growth logs that extract reusable patterns — not diary entries.
Design a goal-oriented agent loop, and review it for the ways loops go wrong — spinning and burning tokens, Goodhart-gaming the verifier, or running a wrong answer to completion. Two actions: (1) WRITE a loop — gate whether to build it, define a machine-decidable goal, pick the loop type, pick a skeleton; (2) REVIEW a loop — run it past five failure modes plus decidability, boundaries, fallback, judge independence, and keep-judgment-with-the-human red lines. Use when designing an autonomous agent loop, or when you already have one and worry it will spin, cheat, or run a wrong answer to the end. Complements the mechanism-layer loop skills (autonomous-loops, continuous-agent-loop) by covering the judgment layer they don't. 中文触发：写 loop、设计 loop、做一个 loop、检查 loop 对不对、loop 体检、loop 会不会跑飞、可判定目标、五个崩法、plan build judge。English triggers: design an agent loop, write a loop, check a loop, loop review, prevent a runaway loop, goal-oriented loop, decidable goal, plan/build/judge.
Stop hook that blocks Claude from finishing until quality checks pass. Detects rationalization patterns (surface text heuristics), stale learning logs (filesystem mtime), and low disk space. Complements self-audit by mechanically enforcing learning capture habits.
React Native and Expo app patterns — Expo Router navigation, state separation (server/client/route/form), TanStack Query data fetching with Zod, performant lists, NativeWind/StyleSheet styling, native APIs, and secure storage. Use when building or editing React Native / Expo screens, components, navigation, or data layers.
Instinct-based learning system that observes sessions via hooks, creates atomic instincts with confidence scoring, and evolves them into skills/commands/agents. v2.1 adds project-scoped instincts to prevent cross-project contamination.
Use this skill when writing new features, fixing bugs, or refactoring code. Enforces test-driven development with 80%+ coverage including unit, integration, and E2E tests.
| name | eval-harness |
| description | Formal evaluation framework for Claude Code sessions implementing eval-driven development (EDD) principles |
| allowed-tools | Read, Write, Edit, Bash, Grep, Glob |
A formal evaluation framework for Claude Code sessions, implementing eval-driven development (EDD) principles.
Eval-Driven Development treats evals as the "unit tests of AI development":
Test if Claude can do something it couldn't before:
[CAPABILITY EVAL: feature-name]
Task: Description of what Claude should accomplish
Success Criteria:
- [ ] Criterion 1
- [ ] Criterion 2
- [ ] Criterion 3
Expected Output: Description of expected result
Ensure changes don't break existing functionality:
[REGRESSION EVAL: feature-name]
Baseline: SHA or checkpoint name
Tests:
- existing-test-1: PASS/FAIL
- existing-test-2: PASS/FAIL
- existing-test-3: PASS/FAIL
Result: X/Y passed (previously Y/Y)
Deterministic checks using code:
# Check if file contains expected pattern
grep -q "export function handleAuth" src/auth.ts && echo "PASS" || echo "FAIL"
# Check if tests pass
npm test -- --testPathPattern="auth" && echo "PASS" || echo "FAIL"
# Check if build succeeds
npm run build && echo "PASS" || echo "FAIL"
Use Claude to evaluate open-ended outputs:
[MODEL GRADER PROMPT]
Evaluate the following code change:
1. Does it solve the stated problem?
2. Is it well-structured?
3. Are edge cases handled?
4. Is error handling appropriate?
Score: 1-5 (1=poor, 5=excellent)
Reasoning: [explanation]
Flag for manual review:
[HUMAN REVIEW REQUIRED]
Change: Description of what changed
Reason: Why human review is needed
Risk Level: LOW/MEDIUM/HIGH
"At least one success in k attempts"
"All k trials succeed"
## EVAL DEFINITION: feature-xyz
### Capability Evals
1. Can create new user account
2. Can validate email format
3. Can hash password securely
### Regression Evals
1. Existing login still works
2. Session management unchanged
3. Logout flow intact
### Success Metrics
- pass@3 > 90% for capability evals
- pass^3 = 100% for regression evals
Write code to pass the defined evals.
# Run capability evals
[Run each capability eval, record PASS/FAIL]
# Run regression evals
npm test -- --testPathPattern="existing"
# Generate report
EVAL REPORT: feature-xyz
========================
Capability Evals:
create-user: PASS (pass@1)
validate-email: PASS (pass@2)
hash-password: PASS (pass@1)
Overall: 3/3 passed
Regression Evals:
login-flow: PASS
session-mgmt: PASS
logout-flow: PASS
Overall: 3/3 passed
Metrics:
pass@1: 67% (2/3)
pass@3: 100% (3/3)
Status: READY FOR REVIEW
/eval define feature-name
Creates eval definition file at .claude/evals/feature-name.md
/eval check feature-name
Runs current evals and reports status
/eval report feature-name
Generates full eval report
Store evals in project:
.claude/
evals/
feature-xyz.md # Eval definition
feature-xyz.log # Eval run history
baseline.json # Regression baselines
## EVAL: add-authentication
### Phase 1: Define (10 min)
Capability Evals:
- [ ] User can register with email/password
- [ ] User can login with valid credentials
- [ ] Invalid credentials rejected with proper error
- [ ] Sessions persist across page reloads
- [ ] Logout clears session
Regression Evals:
- [ ] Public routes still accessible
- [ ] API responses unchanged
- [ ] Database schema compatible
### Phase 2: Implement (varies)
[Write code]
### Phase 3: Evaluate
Run: /eval check add-authentication
### Phase 4: Report
EVAL REPORT: add-authentication
==============================
Capability: 5/5 passed (pass@3: 100%)
Regression: 3/3 passed (pass^3: 100%)
Status: SHIP IT