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。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/adam-s/agent-spec --skill new-eval命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
Compare two eval runs and report what changed. Reads both runs' events, transcripts, and produced artifacts. Writes a short markdown summary classifying differences as regression, improvement, or neutral.
Generalized recursive iteration loop. Runs parallel sub-agents against a target, scores deterministically, diagnoses instruction gaps, applies fixes, and recurses until the stop condition is met or max depth is reached.
Run an evaluation against an eval with a specific config
基于 SOC 职业分类
正在显示 SKILL.md
| name | new-eval |
| description | Scaffold a new evaluation |
| argument-hint | <name> |
Scaffolds an eval directory with template files. Uses the multi-challenge format by default.
See @.claude/reference/eval-definition.md for the full schema, environment setup conventions, and verify.sh contract.
Ask the user what language the eval targets (Python, TypeScript, or other) to select the right environment pattern.
Create these files in evals/$1/:
EVAL.md:
---
name: $1
description: Describe what this eval tests
model: claude-sonnet-4-6
budget: 2.00
---
challenges/challenge-1/prompt.md: Describe the task. Include the environment hint for the language:
.venv/. Use .venv/bin/python3 to run code."node_modules/."challenges/challenge-1/seeds/: Place seed files here (source code, test scripts, data files, and dependency manifests like requirements.txt or package.json).
challenges/challenge-1/setup.sh — create the environment and install deps:
Python:
#!/bin/bash
python3 -m venv .venv
.venv/bin/pip install -r requirements.txt --quiet
TypeScript:
#!/bin/bash
npm install --silent
challenges/challenge-1/verify.sh — run tests and print RESULT. Only verify.sh prints RESULT:, not the test script.
Python:
#!/bin/bash
[ -d .venv ] || python3 -m venv .venv
.venv/bin/pip install -r requirements.txt --quiet 2>/dev/null
.venv/bin/python3 test.py 2>&1
EXIT=$?
if [ $EXIT -eq 0 ]; then
echo "RESULT: PASS"
else
echo "RESULT: FAIL"
fi
TypeScript:
#!/bin/bash
[ -d node_modules ] || npm install --silent 2>/dev/null
node test.js 2>&1
EXIT=$?
if [ $EXIT -eq 0 ]; then
echo "RESULT: PASS"
else
echo "RESULT: FAIL"
fi
configs/baseline/CLAUDE.md:
A coding project.
configs/baseline/settings.json:
{
"permissions": {
"deny": []
}
}
Make scripts executable: chmod +x evals/$1/challenges/challenge-1/verify.sh evals/$1/challenges/challenge-1/setup.sh
Print: "Eval '$1' created. Edit the challenge prompt and seeds, then run /run-eval $1"