chughtapan

Cc Judge — AI skill for Claude Code

AI community

Shared LLM-evals framework for running Claude Code agents in containers, scoring outcomes with an LLM judge, and producing reports.

How to install Cc Judge

This entry records only its repository, not the path inside it, so there is no exact command to give. Open chughtapan/cc-judge and copy the folder into ~/.claude/skills/, or the file into ~/.claude/agents/.

What Cc Judge does

Shared LLM-evals framework for running Claude Code agents in containers, scoring outcomes with an LLM judge, and producing reports. Extracted from moltzap's packages/evals so any project can depend on it.

Alternatives in AI

System Prompts Leaks — Extracted system prompts from ChatGPT (GPT-5.4, GPT-5.3, Codex), Claude (Opus 4.6, Sonnet 4.6, Claude Code), G 38.6k ★
Codex Plugin — 's /codex:adversarial-review command 9.6k ★
Contributing To Claude Skills Library — Thank you for your interest in contributing to the Claude Skills Library 5.3k ★

README

cc-judge

TypeScript-first CLI + SDK for planned Claude Code harness runs, LLM bundle judging, and `summary.md` + `results.jsonl` reports. Telemetry can fan out to Braintrust and Promptfoo through pluggable emitters.

Prerequisites

Node.js 20.11+
pnpm 9+
Claude judge auth through claude auth login or ANTHROPIC_API_KEY=...
Docker, if your harness launches Docker workloads

When `--judge-backend anthropic` is active and `ANTHROPIC_API_KEY` is not set, `cc-judge` runs `claude auth status` before `run`. Successful checks are cached for 24 hours under the user cache directory.

Install

pnpm add cc-judge

Quickstart

Create a harness-backed plan:

project: moltzap
scenarioId: EVAL-005
name: Cold outreach response quality
description: Verify the target agent responds helpfully to a first-contact DM.
requirements:
  expectedBehavior: The agent should answer coherently instead of returning an auth or runtime error.
  validationChecks:
    - Response contains non-empty text
    - Response stays on topic
harness:
  module: ../../packages/runtimes/dist/trace-capture-harness.js
  payload:
    runtime:
      kind: openclaw
    conversation:
      kind: direct
      setupMessage: Hello, can you explain how MoltZap conversations work?

Run it:

cc-judge run ./plans/**/*.yaml --results ./eval-results --log-level info

Successful and failed runs both emit:

eval-results/
  summary.md
  results.jsonl
  details/
    ..yaml

Harness Modules

The plan's `harness.module` path resolves relative to the plan file. The module must export `load(args)` as its default export unless the plan sets `harness.export`.

Minimal module shape:

import { Effect } from "effect";

export default {
  load(args) {
    return Effect.succeed({
      plan: {
        project: args.plan.project,
        scenarioId: args.plan.scenarioId,
        name: args.plan.name,
        description: args.plan.descripti