delegating-work Skill lifts subagent checks to 18/24 from 3/24

The Skill targets a common failure in which a subagent returns 'all done, tests passed' after quietly skipping a test. Its instructions split delegation into four decisions: whether to hand off work, which model tier fits the task, what the request must contain, and how to receive the report. For example, the agent should try two or three tasks before scaling up and should ban deleting, skipping, or loosening tests in the request.

The test used fictional scenarios and role-played subagents, so it measured whether the parent agent wrote prohibitions and verified reports, not whether real subagents would skip tests on their own. The results show clear gains in the file-replacement and lint scenarios, but little difference in the read-only investigation scenario, a gap the author acknowledges.

Hitting the 10-turn limit in several runs suggests verification adds back-and-forth and time; the author advises expecting more turns in real work. The published version also includes rules added after the test that were not measured.