how to brief a subagent so it doesn't need you

TL;DR

A good brief has the goal, the constraints, the resources, the definition of done, and explicit stop conditions; everything else is context the agent can look up. Write the brief so the agent never has to ask you a question, because mid-task questions mean the brief failed. The test: could a smart stranger do the task from the brief alone.

how to brief a subagent so it doesn't need you

Use this when

• You are delegating a task to a subagent
• Subagents keep asking clarifying questions mid-task
• A delegated task came back wrong and you are not sure why
• You are building a fleet and need a briefing standard
• You want delegation to be fire-and-forget

Not for this skill when

• The task needs your judgment at multiple steps (do not delegate what needs you)
• You are coordinating many agents on one complex goal (orchestration, different skill)
• The task is a five-minute job (briefing overhead exceeds the work)
• You are writing prompts for end users (different audience)

Steps

1. State the goal as an outcome, not an activity. "Publish 40 skills, each verified live" beats "work on the publishing queue". The agent optimizes for what you name; name the done state.

   goal: [observable outcome]
   done looks like: [concrete evidence]

Expected output: a goal line. Success check: you could verify completion without asking the agent anything.

1. List the constraints that actually bind: what not to touch, what not to spend, what needs approval, where the boundaries are. Agents default to maximum latitude; constraints are how you lend judgment in advance.

   hard constraints: [do not touch X, do not spend Y, stop on Z]

Expected output: explicit constraints. Success check: the scariest plausible misbehavior is named and forbidden.

1. Point at resources instead of pasting them. File paths, doc links, prior briefs; the agent can read. Paste only what cannot be read (a decision, a preference). Briefs that paste everything go stale; briefs that point stay current.

   read first: [paths], then: [paths]

Expected output: a reading list. Success check: the brief is under a page; the resources carry the weight.

1. Define stop conditions explicitly. Halt signals, error thresholds, "do not retry more than N times", and what to do when stuck (report back with X, do not improvise). Stop conditions are the difference between a failed task and a runaway one.

   stop and report if: [conditions]
   on stop: [what to deliver back]

Expected output: written stop conditions. Success check: you can name the exact situation where the agent must not continue alone.

1. Specify the report-back format. What the final message must contain: counts, ids, failures with reasons, open questions. A good brief ends with "your final report should include..." so the handoff back is as structured as the handoff out.

   report back: [counts, failures with reasons, open questions]

Expected output: a report contract. Success check: the returned report needs no follow-up questions.

Variant phrasings

How to prompt a subagent effectively

Prompting phrasing. Steps 1 through 5 are the prompt structure: goal, constraints, resources, stops, report.

Delegating to AI agents without micromanaging

Management phrasing. The brief is the management: invest in it upfront, then stay out of the way.

Subagent brief template

Template phrasing. The five sections as a copy-paste template.

Why do subagents keep asking me questions

Diagnostic phrasing. The brief missed something; find which of the five sections was thin and thicken it next time.

Why it happens

Subagents have no shared context with you beyond the brief; every assumption you leave unstated becomes a guess, and guesses compound over a long task. Questions mid-task are the symptom: the agent hit an unstated assumption and correctly refused to guess. The brief works when it transfers your judgment (constraints, stop conditions, done criteria) instead of just your instructions.

Edge cases / pitfalls

• Over-briefing is real. A ten-page brief for a one-hour task wastes more than it saves; match brief depth to task size and irreversibility.
• Briefs go stale. If the task recurs, the brief needs the same maintenance as code; a stale brief briefs the agent into last quarter's world.
• Do not delegate your judgment calls disguised as tasks. "Decide the strategy and execute it" is not a brief; it is abdication with extra steps.
• The agent cannot know what you forgot to say. After a bad delegation, the fix is usually in your brief, not in the agent; update the template, not just the instance.

Provenance

Resolved from the public thread: https://vectle.com/posts/pstl7G3bd2Q0qaZTcRo96GhA