Skillgrade

May 18, 2026 · View on GitHub

The easiest way to evaluate your Agent Skills. Tests that AI agents correctly discover and use your skills.

See examples/ — superlint (simple) and angular-modern (TypeScript grader).

Browser Preview

Quick Start

Prerequisites: Node.js 20+, Docker

npm i -g skillgrade

1. Initialize — go to your skill directory (must have SKILL.md) and scaffold:

cd my-skill/
GEMINI_API_KEY=your-key skillgrade init    # or ANTHROPIC_API_KEY / OPENAI_API_KEY
# Use --force to overwrite an existing eval.yaml

Generates eval.yaml with AI-powered tasks and graders. Without an API key, creates a well-commented template.

2. Edit — customize eval.yaml for your skill (see eval.yaml Reference).

3. Run:

GEMINI_API_KEY=your-key skillgrade --smoke

The agent is auto-detected from your API key: GEMINI_API_KEY → Gemini, ANTHROPIC_API_KEY → Claude, OPENAI_API_KEY → Codex. Override with --agent=claude.

4. Review:

skillgrade preview          # CLI report
skillgrade preview browser  # web UI → http://localhost:3847

Reports are saved to $TMPDIR/skillgrade/<skill-name>/results/. Override with --output=DIR.

Presets

Flag	Trials	Use Case
`--smoke`	5	Quick capability check
`--reliable`	15	Reliable pass rate estimate
`--regression`	30	High-confidence regression detection

Options

Flag	Description
`--eval=NAME[,NAME]`	Run specific evals by name (comma-separated)
`--grader=TYPE`	Run only graders of a type (`deterministic` or `llm_rubric`)
`--trials=N`	Override trial count
`--parallel=N`	Run trials concurrently
`--agent=gemini\|claude\|codex\|acp\|opencode`	Override agent (default: auto-detect from API key)
`--provider=docker\|local`	Override provider
`--acp-command=CMD`	ACP agent command (e.g., `gemini --acp`)
`--opencode-agent=NAME`	OpenCode agent (build\|plan\|explore)
`--opencode-model=MODEL`	OpenCode model (provider/model format)
`--output=DIR`	Output directory (default: `$TMPDIR/skillgrade`)
`--validate`	Verify graders using reference solutions
`--ci`	CI mode: exit non-zero if below threshold
`--threshold=0.8`	Pass rate threshold for CI mode
`--preview`	Show CLI results after running

eval.yaml Reference

version: "1"

# Optional: explicit path to skill directory (defaults to auto-detecting SKILL.md)
# skill: path/to/my-skill

defaults:
  agent: gemini          # gemini | claude | codex | acp
  provider: docker       # docker | local
  trials: 5
  timeout: 300           # seconds
  threshold: 0.8         # for --ci mode
  grader_model: gemini-3-flash-preview  # default LLM grader model
  grader_provider: gemini               # default LLM grader provider: gemini | anthropic | openai
  acp:                   # ACP agent configuration (optional)
    command: gemini --acp  # command to start ACP-compatible agent
    env:                  # optional environment variables
      DEBUG: "1"
  docker:
    base: node:20-slim
    setup: |             # extra commands run during image build
      apt-get update && apt-get install -y jq
  environment:           # container resource limits
    cpus: 2
    memory_mb: 2048

tasks:
  - name: fix-linting-errors
    instruction: |
      Use the superlint tool to fix coding standard violations in app.js.

    workspace:                           # files copied into the container
      - src: fixtures/broken-app.js
        dest: app.js
      - src: bin/superlint
        dest: /usr/local/bin/superlint
        chmod: "+x"

    graders:
      - type: deterministic
        setup: npm install typescript    # grader-specific deps (optional)
        run: npx ts-node graders/check.ts
        weight: 0.7
      - type: llm_rubric
        rubric: |
          Did the agent follow the check → fix → verify workflow?
        provider: gemini                 # optional: gemini (default) | anthropic | openai
        model: gemini-2.0-flash          # optional model override
        weight: 0.3

    # Per-task overrides (optional)
    agent: claude
    grader_provider: anthropic   # override default LLM grader provider
    trials: 10
    timeout: 600

String values (instruction, rubric, run) support file references — if the value is a valid file path, its contents are read automatically:

instruction: instructions/fix-linting.md
rubric: rubrics/workflow-quality.md

Graders

Deterministic

Runs a command and parses JSON from stdout:

- type: deterministic
  run: bash graders/check.sh
  weight: 0.7

Output format:

{
  "score": 0.67,
  "details": "2/3 checks passed",
  "checks": [
    {"name": "file-created", "passed": true, "message": "Output file exists"},
    {"name": "content-correct", "passed": false, "message": "Missing expected output"}
  ]
}

score (0.0–1.0) and details are required. checks is optional.

Bash example:

#!/bin/bash
passed=0; total=2
c1_pass=false c1_msg="File missing"
c2_pass=false c2_msg="Content wrong"

if test -f output.txt; then
  passed=$((passed + 1)); c1_pass=true; c1_msg="File exists"
fi
if grep -q "expected" output.txt 2>/dev/null; then
  passed=$((passed + 1)); c2_pass=true; c2_msg="Content correct"
fi

score=$(awk "BEGIN {printf \"%.2f\", $passed/$total}")
echo "{\"score\":$score,\"details\":\"$passed/$total passed\",\"checks\":[{\"name\":\"file\",\"passed\":$c1_pass,\"message\":\"$c1_msg\"},{\"name\":\"content\",\"passed\":$c2_pass,\"message\":\"$c2_msg\"}]}"

Use awk for arithmetic — bc is not available in node:20-slim.

LLM Rubric

Evaluates the agent's session transcript against qualitative criteria:

- type: llm_rubric
  rubric: |
    Workflow Compliance (0-0.5):
    - Did the agent follow the mandatory 3-step workflow?

    Efficiency (0-0.5):
    - Completed in ≤5 commands?
  weight: 0.3
  provider: gemini           # gemini (default) | anthropic | openai
  model: gemini-2.0-flash    # optional, auto-detected from API key

The provider field selects which LLM API to call:

Provider	API Key Env Var	Base URL Env Var (optional)	Default Model
`gemini`	`GEMINI_API_KEY`	-	`gemini-3-flash-preview`
`anthropic`	`ANTHROPIC_API_KEY`	`ANTHROPIC_BASE_URL`	`claude-sonnet-4-20250514`
`openai`	`OPENAI_API_KEY`	`OPENAI_BASE_URL`	`gpt-4o`

ANTHROPIC_BASE_URL and OPENAI_BASE_URL enable custom/self-hosted endpoints (Ollama, vLLM, etc.).

Combining Graders

graders:
  - type: deterministic
    run: bash graders/check.sh
    weight: 0.7      # 70% — did it work?
  - type: llm_rubric
    rubric: rubrics/quality.md
    weight: 0.3      # 30% — was the approach good?

Final reward = Σ (grader_score × weight) / Σ weight

CI Integration

Use --provider=local in CI — the runner is already an ephemeral sandbox, so Docker adds overhead without benefit.

# .github/workflows/skillgrade.yml
- run: |
    npm i -g skillgrade
    cd skills/superlint
    GEMINI_API_KEY=${{ secrets.GEMINI_API_KEY }} skillgrade --regression --ci --provider=local

Exits with code 1 if pass rate falls below --threshold (default: 0.8).

Tip: Use docker (the default) for local development to protect your machine. In CI, local is faster and simpler.

Environment Variables

Variable	Used by
`GEMINI_API_KEY`	Agent execution, LLM grading (`provider: gemini`), `skillgrade init`
`ANTHROPIC_API_KEY`	Agent execution, LLM grading (`provider: anthropic`), `skillgrade init`
`OPENAI_API_KEY`	Agent execution (Codex), LLM grading (`provider: openai`), `skillgrade init`
`ANTHROPIC_BASE_URL`	LLM grading (`provider: anthropic`) — custom Anthropic-compatible endpoint
`OPENAI_BASE_URL`	LLM grading (`provider: openai`) — custom OpenAI-compatible endpoint (Ollama, vLLM, etc.)

Variables are also loaded from .env in the skill directory. Shell values override .env. All values are redacted from persisted session logs.

OpenCode Agent

OpenCode is an AI coding agent that supports multiple AI models and specialized subagents.

Quick Start

# Use OpenCode with default agent and model
skillgrade --agent=opencode

# Specify OpenCode agent (build|plan|explore)
skillgrade --agent=opencode --opencode-agent=build

# Specify both agent and model (provider/model format)
skillgrade --agent=opencode --opencode-agent=build --opencode-model=anthropic/claude-sonnet-4-20250514

OpenCode Agents

Agent	Description
`build`	Default primary agent with full tool access
`plan`	Read-only planning/analysis agent
`explore`	Fast codebase exploration agent

OpenCode Models

Models are specified in provider/model format:

Model	Format
Claude Sonnet 4	`anthropic/claude-sonnet-4-20250514`
GPT 5.1 Codex	`opencode/gpt-5.1-codex`

CLI Options

Flag	Description
`--agent=opencode`	Use OpenCode agent
`--opencode-agent=NAME`	OpenCode agent (build\|plan\|explore)
`--opencode-model=MODEL`	OpenCode model (provider/model format)

How It Works

skillgrade invokes OpenCode CLI with opencode run
Passes instruction via temp file to avoid shell escaping issues
Supports both agent and model specification
Works with --provider=docker or --provider=local

Agent Client Protocol (ACP) is an open protocol that standardizes communication between AI coding agents and clients. Using an ACP-compatible agent allows you to evaluate skills without managing API keys directly.

Quick Start

# Use Gemini CLI in ACP mode (requires gemini CLI installed)
skillgrade --agent=acp --acp-command="gemini --acp"

# Or configure in eval.yaml

defaults:
  agent: acp
  acp:
    command: gemini --acp

ACP-Compatible Agents

Any agent that supports the ACP protocol can be used:

Agent	Command
Gemini CLI	`gemini --acp`
Other ACP agents	Check agent documentation

How It Works

skillgrade starts the ACP agent as a subprocess
Communication happens via JSON-RPC 2.0 over stdio
No API key required — authentication is handled by the ACP agent
Works best with --provider=local since the ACP agent needs to be available in your environment

CLI Options

Flag	Description
`--agent=acp`	Use ACP-compatible agent
`--acp-command=CMD`	Command to start the ACP agent

The --acp-command can also be set in eval.yaml under defaults.acp.command.

Best Practices

Grade outcomes, not steps. Check that the file was fixed, not that the agent ran a specific command.
Instructions must name output files. If the grader checks for output.html, the instruction must tell the agent to save as output.html.
Validate graders first. Use --validate with a reference solution before running real evals.
Start small. 3–5 well-designed tasks beat 50 noisy ones.

For a comprehensive guide on writing high-quality skills, check out skills-best-practices. You can also install the skill creator skill to help author skills:

npx skills add mgechev/skills-best-practices

License

MIT

Inspired by SkillsBench and Demystifying Evals for AI Agents.