GitHub radar
Unlazy: Stop AI Agents from Calling Done Too Soon
A SKILL.md-based completion discipline for Claude Code: agents define acceptance criteria, execute, and verify — before they're allowed to report done.
Unlazy is a completion discipline framework for AI agents that enforces structured verification through runnable gates and approval records. It uses an 'acceptance ledger first' methodology: write the expected outcomes before execution, then verify results against those criteria. Gate contracts are Markdown files combining a shell command (CHECK:), expected output (EXPECT:), and evidence tracking. Approval records are stored locally with bindings to specific gate definitions. An optional Stop hook for Claude Code blocks the agent from finishing until all gates pass. Supports parallel execution via --jobs flag (1-64 concurrent). Reached 3,000 stars in its first month after publishing — one of the fastest-growing skills in the Claude Code ecosystem.
Why a vibe-coder should care
If you've been burned by agents that declare success before the job is really done — broken tests, missing files, half-finished refactors — unlazy makes premature 'done' impossible. It slows the agent down by seconds at the end but saves you the time of catching and redoing incomplete work.
How to install
Copy this and send it to your agent — Claude Code, Codex, any of them:
Install the unlazy skill from https://github.com/Leonxlnx/unlazy — follow the README, install via `npx skills add Leonxlnx/unlazy`, wire up the Stop hook for Claude Code, and show me it works: create a simple gate contract and run the check.
Runs on a regular laptop — you'll need Node.js and Python 3.9+.
Open on GitHub▌ More finds