response-rater

oimiragieo 的头像作者 oimiragieo

运行无头AI CLI(Claude Code `claude`、Google `gemini`,可选GitHub `copilot`)对助手响应进行评分,并返回可操作反馈及改进后的响应;用于响应质量审计、提示/文档审查以及“让另一个AI评估此答案”的工作流程。

📖 中文导读:运行无头AI CLI(Claude Code `claude`、Google `gemini`,可选GitHub `copilot`)对助手响应进行评分,并返回可操作反馈及改进后的响应;用于响应质量审计、提示/文档审查以及“让另一个AI评估此答案”的工作流程。 以下正文为该技能的原始 SKILL.md 内容(英文),上方安装命令可直接使用。

name: response-rater description: Run headless AI CLIs (Claude Code claude, Google gemini, optionally GitHub copilot) to rate an assistant response against a rubric and return actionable feedback plus a rewritten improved response; use for response quality audits, prompt/docs reviews, and "have another AI critique this answer" workflows. version: 1.0.0 model: opus allowed-tools: bash

Response Rater

Use this skill to get independent critiques of an assistant response by calling external AI CLIs in headless mode and aggregating the results.

Quick start

  1. Save the response you want reviewed to a file (or copy/paste it when prompted).

  2. Ensure you have at least one provider authenticated:

  • Claude Code: either (a) you’re already logged in via subscription/session, or (b) set ANTHROPIC_API_KEY
  • Gemini CLI: either (a) you’re already logged in via subscription/session, or (b) set GEMINI_API_KEY or GOOGLE_API_KEY
  • Copilot CLI: sign in with the CLI’s normal auth flow (varies by setup)
  1. Run:

node codex-skills/response-rater/scripts/rate.js --response-file <path> --providers "claude,gemini"

Or via stdin:

Get-Content response.txt | node codex-skills/response-rater/scripts/rate.js --providers "claude,gemini"

Vocabulary review (Privacy Airlock Stage 1)

Get-Content server/models/sentencepiece/vocabulary.json | node codex-skills/response-rater/scripts/rate.js --providers "claude,gemini" --template vocab-review

If Claude times out, increase timeout:

Get-Content server/models/sentencepiece/vocabulary.json | node codex-skills/response-rater/scripts/rate.js --providers "claude,gemini" --template vocab-review --timeout-ms 240000

Auth behavior

By default the runner uses --auth-mode session-first:

  1. Try the CLI using your existing logged-in session/subscription
  2. If that fails and env keys exist, retry using env keys

To flip the order:

node codex-skills/response-rater/scripts/rate.js --response-file response.txt --auth-mode env-first --providers claude,gemini

Direct headless commands (no script)

Claude:

Get-Content response.txt | claude -p --output-format json --permission-mode bypassPermissions

Gemini:

Get-Content response.txt | gemini --output-format json --model gemini-2.5-flash

What it outputs

JSON to stdout:

  • per-provider raw output
  • parsed scores/summary/improvements/rewrite when the model returns valid JSON

Notes / constraints

  • This skill requires network access to contact the providers.
  • If a provider is missing credentials, it will be skipped with a clear error.
  • Keep the reviewed response reasonably sized; start with the exact section you want critiqued.

Install

Copy codex-skills/response-rater to your Codex skills directory (usually $CODEX_HOME/skills/response-rater).