Unlazy, an agent skill against model laziness, passed 2,986 stars

An MIT-licensed agent skill called unlazy, aimed at reducing laziness and premature task completion in coding agents, reached 2,986 GitHub stars over its 30-day window. The project targets Claude Code and Codex CLI and works with any agent that can consume a Markdown skill file.

The Depth Tree method, in one paragraph

The idea it packages is a task-decomposition pattern the author calls a Depth Tree. A task is split N layers deep, and every leaf inherits the full time budget of the whole task rather than a share of it. The author's argument is that this makes effort multiply with depth, so a model that would otherwise stop at a shallow, plausible answer keeps going through the sub-questions the tree exposed. The skill file is the enforcement mechanism, not a runtime: the agent reads it and the pattern shapes what the agent does next.

The research the project cites

The README bases the design on a list of dated papers. It lists Test-Time Scaling in Reasoning Models (August 2026, COLM 2026), Measuring AI Ability to Complete Long Software Tasks (July 2026, NeurIPS 2025), SlopCodeBench (May 2026), LLM Pitfalls in Software Engineering and When More Thinking Hurts (both April 2026), METR Time Horizon 1.1 (January 2026), Quantifying Laziness (December 2025, arXiv 2512.20662), OptimalThinkingBench (October 2025), s1: Simple test-time scaling (March 2025), and Thoughts Are All Over the Place (February 2025).

That list is the interesting part. It is a real bibliography for a genuine failure mode, and it is what makes the project more than a prompt tip. That research covers two problems: models that stop early on hard tasks (laziness, underthinking) and models that waste effort on easy ones (overthinking). A skill that always maximises depth will trigger the second problem on tasks that did not need it, so this is worth reading alongside the README rather than instead of it.

What the star count measures

Stars are attention. The 2,986 number over 30 days is genuine reach at the scale where a project is worth looking at, and the traction on this class of skill files has been high through August. Attention is a signal about reach, not a benchmark on output quality; the project includes no benchmark run, so any claim about better code from the method would need one.

If your team is running long agentic tasks, or hitting model runs that stop at the first plausible answer, the Depth Tree is a cheap thing to try. The install is a skill file: a copy into the agent's skill directory and a mention in the prompt. The cost of getting it wrong is a small number of extra tokens on tasks that did not need the extra depth. The cost of getting it right is fewer of the shallow-answer failures that waste hours.

Measure it yourself. Pick two tasks the team has actually failed on in the last week, one truly hard and one that only looked hard. Run each with the skill and without it, and keep the outputs. If the pattern helps on the first and does not waste time on the second, it is worth keeping. If it slows the easy one down without changing the hard one, it does not.

Source

Primary source: Leonxlnx/unlazy on GitHub.