Delegation applies before you spawn a subagent or dispatch a build, review or research task, and again before you trust its report. It treats a delegate as a partner who knows only what you hand it, so the brief carries the goal, the binding constraints and how the result will be proved.
It also lists what never delegates, when handing off is worth it, how to choose among Claude subagents and headless CLIs, how to run several agents at once, and how to verify what comes back.
What You Get
- A three-part brief method: the goal and decision, the binding constraints and traps, and the proof
- A list of what never delegates, with the reasoning
- A table comparing delegates: Claude subagents, Codex, Grok, headless Claude Code and the Antigravity CLI
- Rules for model and effort on Claude subagent spawns, and for parallel work with separate worktrees
- A checklist for when a delegate returns: read the diff, check the disk, rerun tests, compare the test inventory
- A brief template with a handoff section
Common Use Cases
- Writing the brief for a build stage handed to another agent
- Deciding whether a task is worth delegating at all
- Running two builds in parallel without them overwriting each other
- Checking a delegate’s claim that tests pass
What Makes It Unique
Most delegation advice stops at writing a prompt. This skill covers the return trip: a delegated diff is unreviewed code, a round with no write access can return a convincing report with no changes, and a builder’s passing report is a claim until you rerun the tests. It also sets the model explicitly on every spawn, since inheriting the parent’s model is the absence of a routing decision.
Under The Hood
The skill is instructions and one reference template, with no scripts. It shapes how an agent writes briefs and checks results. Along the way it tells the agent to run git commands to inspect status and ranges, to use separate git worktrees for concurrent writers, to run test suites to verify a handoff, and to read agent transcripts with a grep command to check the effort a subagent used. It does not fetch web pages or call APIs.