The Unlazy Skill Stops Claude Faking Finished Work

Sometimes an AI agent tells you a job is done when part of it never happened. Researchers call this partial compliance: the model does some of a multi-part request and stops early. Unlazy is a free, open-source skill that helps with this. It makes Claude check its work against clear tests before it's allowed to say a task is finished.

A few plain terms first. Claude Code is Anthropic's coding tool. A skill is a set of instructions Claude follows for one kind of job. A gate is a check with a command to run and the result it should give.

1. Install the skill

Unlazy lives on GitHub at Leonxlnx/unlazy. Use the original repo, not one of the copies. The easiest way to install it is with the skills tool. In your terminal, run:

npx skills add Leonxlnx/unlazy

Add -g to install it for your whole user account. You can also clone the repo into ~/.claude/skills/unlazy by hand. The checker needs Node 16 or newer.

Unlazy on GitHub

2. Give it a job

Start your request with /unlazy so Claude knows to use the skill. For example:

/unlazy tree 3 build a simple page that lists the books I read each week, and check every part works before you tell me it's done

3. Let it split the work into small jobs

Unlazy breaks your project into a tree of smaller tasks. Each one gets a short checklist and a line saying what finished actually looks like.

4. Check the gates file

It writes a gates file. Every item has a command to run (CHECK), the result it should give (EXPECT) and a line for the proof (EVIDENCE). A gate only passes when the command really gives that result, so Claude can't just say it's done. Gate commands are real commands, so read them before you approve them.

5. Choose how deep it goes

The number after tree sets how many layers deep it splits the work. Use a small number for a small job and a bigger number for a big project.

6. Run tasks side by side with worktrees

By default, Unlazy works through tasks one after another, which can take a while on a big project. Git worktrees let several Claude agents work at the same time, each in its own copy of the project folder. The free parallel-worktrees skill sets this up for you. Then you can ask Claude:

You can run several agents at the same time. Give each independent task its own git worktree, run them in parallel, then check every gate before you merge the results back together.

Parallel worktrees skill on GitHub

Good to know

• Unlazy is a third-party skill, not made by Anthropic.
• It can only check what your gates test, so make each check look at the real result.
• Claude Code needs a paid Claude plan, like Pro or Max.