Lean Skills

Official Agent Skills for developing with Lean 4.

These skills help AI coding agents work effectively with Lean 4 code — writing proofs, setting up development environments, debugging toolchain regressions, and following project conventions.

Skills included

Installation

Claude Code

In Claude Code, type /plugin, under "Marketplaces" add "https://github.com/leanprover/skills.git", then under "Plugins" install the lean plugin.

Or in a terminal

claude plugin marketplace add https://github.com/leanprover/skills.git
claude plugin install lean@leanprover

Gemini CLI

gemini skills install https://github.com/leanprover/skills --path skills

Codex CLI

Inside a Codex session:

$skill-installer install https://github.com/leanprover/skills/tree/main/skills/*

Or clone and copy the skills:

git clone https://github.com/leanprover/skills.git /tmp/lean-skills
cp -r /tmp/lean-skills/skills/* ~/.codex/skills/

Other tools

Clone this repository and point your tool at the skills/ directory.

Testing

Every skill should be validated by test cases that demonstrate it actually improves agent performance. As base model capabilities increase over time, skills that no longer provide value can be identified and removed.

Each test case is a YAML file in skills/<name>/tests/ specifying a git repo, commit SHA, and prompt. Test infrastructure and results live in a separate repository: leanprover/skills-testing.

Checking validation status

# Check that all skills have current satisfactory results
scripts/check-validation

# Check a specific skill
scripts/check-validation lean-proof

This is also run in CI on every push and pull request.

Running tests

Clone the testing repo into this directory (it's gitignored):

git clone https://github.com/leanprover/skills-testing.git

Then run tests:

# Run all tests for a skill
skills-testing/scripts/run-skill-tests lean-proof

# Judge completed runs
skills-testing/scripts/judge-all

# View results
skills-testing/scripts/summary --latest

Commit results back to skills-testing so CI passes.

Test case format

repo: "https://github.com/owner/repo.git"
sha: "abc123"
prompt: |
  Your prompt here.
description: "What this tests"
# Optional: subdirectory, claude_flags, timeout (default 600s)

Judge verdicts

• satisfactory — the skill meaningfully improved the response
• not_needed — the base model performed equally well without the skill
• needs_improvement — the skill helped but could be better (suggestions provided)

License

Apache-2.0