---
id: prj-delegation-brief-for-an-agent
schema: project/1
layer: garden
title: Write a handover brief before you give an agent a real task
outcome: One task handed to an AI agent with a written brief, a decision about whether it was worth handing over, and a check of the result against the "done" line you wrote in advance.
roles: [team lead, manager, programme coordinator, analyst, anyone starting to use AI agents]
time_box: 45 minutes, plus the agent's run time
tool_access: An AI tool that can work on a multi-step task by itself, for example Claude Cowork, ChatGPT agent mode or a similar agent. A regular chat assistant works for a smaller version. Only give it files and accounts your organisation allows.
practises: [cmp-describe-the-task, cmp-steering, cmp-check-the-result]
concepts: [con-delegating-to-ai, con-agents, con-steering, con-human-in-the-loop, con-verification, con-jagged-frontier]
skills: [build-brief, prove-it]
resources: [res-other-mollick-management-as-ai-superpower, res-other-mollick-working-with-wizards, res-other-mollick-which-ai-summer-2026]
builds_on: [prj-plan-with-clarifying-questions]
source:
  url: https://www.oneusefulthing.org/p/management-as-ai-superpower
  claim: Handing work to an AI agent uses the same skills as handing it to a person. A short brief that states the goal and its purpose, the limits of the agent's authority, what done looks like, the outputs, checkpoints and a self-check gets better results. Whether a task is worth handing over depends on how long it takes you, how often the AI succeeds and how long asking and checking take.
  evidence: Ethan Mollick's January 2026 essay, which draws on his teaching (executive MBA students building prototypes in four days) and a worked estimate from OpenAI's GDPval benchmark. The brief checklist is his advice from experience, not a tested intervention. The steps, sample task and checks on this card are the commons' own.
  limits: The time-saving estimate is an average over benchmark tasks, and your task's success rate can be far lower. One run tells you about one task in one tool. The card doesn't show that a brief improves every agent; it gives you a way to find out for your own work.
origin:
  kind: ai-proposed
  by: "seed agent 2026-10-16"
  rationale: Turns Mollick's delegation checklist into a hands-on exercise with a worth-it estimate before the run and a done-check after it.
---

## Why this matters

Agents can now take a task, work on it for many minutes and come back with a finished file. That makes the handover the step that matters most. A vague request gets a confident result that answers a different question, and you only find out after you've read all of it. Managers already know how to brief a new colleague. The same brief works for an agent, and writing it first also tells you whether the task was worth handing over at all.

## Smallest useful build

A one-page brief for one real task, a quick worth-it estimate, the agent's output, and a short note scoring the output against your "done" line. About 45 minutes of your time.

## Sample inputs

Use a task from your own work that would take you one to three hours. If you want a practice task, use this invented one:

```text
Task: Compare the volunteer sign-up pages of three local food banks (use any three public websites you like) and recommend two changes to our own sign-up page.
Our page asks for: name, email, phone, date of birth, emergency contact, t-shirt size, how you heard about us. It has no mobile layout.
```

## Steps

1. **Decide whether to hand it over.** Before writing anything, jot down three numbers: how long the task takes you, how many tries out of ten you think the AI would get right, and how long writing the brief and checking the result will take. If asking and checking would take longer than doing it, do it yourself or pick a bigger task. Then ask for a second opinion:

   ```text
   I'm deciding whether to hand a task to an AI agent. It takes me about [time]. I think the agent would get it right [x] times in 10. Writing the brief and checking will take about [time]. The task: [paste]. Which part of this task is most likely to go wrong, and would you change my estimate? Don't do the task.
   ```

2. **Write the brief yourself.** Use six short headings: *Goal and why*, *What you may and may not do* (for example, "read public pages only; don't fill in any forms"), *Done looks like*, *What to hand back*, *Check in with me when*, and *Before you report back, check that*. Write every line in your own words. Then ask the AI to find gaps without filling them:

   ```text
   Below is a brief I'm about to give an AI agent. List any place where a capable new colleague would have to guess what I meant. Don't rewrite the brief or answer the questions yourself. Brief: [paste]
   ```

3. **Run it and stay reachable.** Paste the brief into the agent. Keep approvals on for anything that sends, buys, deletes or submits. When it checks in, answer briefly and note what it asked; those questions show what your brief left out.

4. **Score the result against your "done" line.** Before reading the output closely, reread your *Done looks like* section. Mark each line met, partly met or missed, pointing to the place in the output. Then write one sentence: "Handing this over saved / cost me about ___ minutes, because ___."

## Check it

- Every "met" mark points to a specific part of the output.
- Anything the agent claims to have read or checked (a page, a figure, a date) is spot-checked by you on at least two items.
- Your one-sentence verdict uses real minutes, compared with the estimate from step 1.

## Break it

Run the same task again with only the first line of your brief, the goal and nothing else. Compare the two outputs. If the short version is just as good, this kind of task doesn't need a full brief and you can save the effort. If it drifts, note exactly which heading it needed most.

## Reflect

- Which heading was hardest to write? That's usually where your own thinking about the task was still vague.
- Did the agent check in where you expected, or somewhere else?
- Would you give this brief to a new human colleague as it stands?

## Next iteration

Save the brief as a template for this kind of task. Next time, hand over a slightly bigger task, or one where checking is harder, and keep the same scoring habit. When a brief works three times, consider turning it into a reusable instruction with `make-it-repeatable`.