Orchestrating Agent Delegation

Use when about to dispatch a subagent (or multiple) to execute work, when writing a prompt for another agent that will do the coding, when planning parallel work across agents, when structuring an adversarial review panel, or when a subagent reports "done" and you're about to trust its summary instead of the diff. Names the CTO-as-orchestrator model, the model-tiering table (haiku for exploration, sonnet for build + review + scoped fixes, opus for design agents and load-bearing architecture), the trust-but-verify discipline (verify every "done" against the diff, don't trust the agent's summary), refute-first adversarial reviewer prompts, the design→rule→build separation that prevents agents from silently picking architecture, and the "never delegate the 'should we?' question" rule. Cite when a prompt says "decide whether to do X" (smell — the orchestrator should decide already), when a review panel is prompted to grade rather than refute, when a single agent is asked to both design and build a non-trivial sli

thrillmade Updated 3 repo stars

File contents

thrillmade/agent-skills commit 32d2bae0bd

Frequently asked questions

npx get-skill add thrillmade/orchestrating-agent-delegation