It explains Unlazy, a GitHub skill that forces AI agents to prove task completion by writing a plan and a gates ledger. The skill breaks large tasks into a tree of sub-tasks, assigns checkable commands and evidence, and runs verifications to prevent agents from claiming unfinished work. The video also identifies a speed issue and shows parallelization fixes.
– Problem: Agents get “lazy” as context windows grow — they stop early, omit hard parts, or claim completion without proof.
– Solution (Unlazy): Break tasks into a tree, write plan.md and gates.md, require a command, expected output, and recorded evidence for every gate; a checker verifies results.
– Modes & usage: Depth controls breakdown; solo mode (<=3) keeps one session, orchestrated mode (>=4) uses files and fresh sub-agents; installs into .agents/.claude.
– Performance fix: Original implementation serialized sub-agents, causing multi-hour runs; modifying the skill to run sub-agents in parallel (and optionally route tasks to appropriate models) dramatically reduced runtime.
Quotes:
It doesn’t tell you the agent is done, it proves it.
A tick box marked ‘pending’ is worse than an empty box.
Each task should be worth at least 10 minutes of real work.
Statistics
| Upload date: | 2026-08-20 |
|---|---|
| Likes: | 2271 |
| Comments: | 97 |
| Fan Rate: | 1.41% |
| Statistics updated: | 2026-09-11 |
