Give an AI agent a clear brief and sensible permissions

Translate a broad goal into permitted actions, review points, and evidence of completion.

“You may draft” and “you may send” are different permissions.

Name the workspace and the output

An agent needs to know where to work and what to produce. “Clean up my business” is too broad to evaluate. “Create a draft FAQ from these approved policy documents” gives the work an object. Specify which sources it may use, where the result belongs, and what it should do when information is missing. Keep credentials and unrelated private material outside the working set.

Separate preparation from external action

Reading notes, preparing a draft, and proposing a file organization are different from sending an email, publishing a page, or deleting files. State which actions are allowed and which require your review. Use the application’s permission controls too; wording in a prompt is not an access-control system. A good brief and sensible technical permissions support each other.

Design a useful checkpoint

A checkpoint should show you something concrete. “Ask me before proceeding” can generate vague interruptions. “Show the first five classified records, the category rules, and ambiguous cases before processing the rest” lets you correct the interpretation. Choose checkpoints where an error would be costly to undo, not after every trivial step.

Ask for a receipt you can inspect

At completion, request the artifact, a concise description of changes, checks performed, unresolved items, and any actions still awaiting approval. Then inspect what matters. A file cleanup proposal should account for every file. A draft message should contain the right recipient details and commitments before it is sent. Keep the original material or a recoverable version while learning the workflow.

A worked example

Original teaching example, not a recorded model run.

Vague starting point: Do whatever it takes to finish.

A more useful brief

Prepare a draft response using the attached policy. You may read these files and create a draft in this folder. Do not send messages or modify the source files. Flag any promise the policy does not support. Return the draft and the passages used.

Why this helps: The agent has room to do useful preparation while the consequential action has a clear review point.

Your turn

Write an agent brief with five lines: goal, inputs, permitted actions, review point, and completion evidence.

You are looking for this: You can distinguish the work the agent may complete autonomously from the action you intend to review.

A quick check

Which is a technical permission control?

1. Writing “be careful” in the prompt
2. Restricting the tool to the folder and actions it needs
3. Calling the assistant an expert

Answer: Restricting the tool to the folder and actions it needs

Prompt instructions communicate intent; tool permissions determine actual access.

Sources and how to read this guide

Examples and exercises are original teaching material, not recorded model runs, measured time savings, or certifications. Official sources were consulted September 6, 2026. Product behavior can change.

• Anthropic · Building effective agents: Distinguishes predefined workflows from systems that choose their own steps.

Take the idea further

• Chat, workflow, or agent: choose how much work to delegate
• Keep a long AI project on track with a useful handoff

All learning paths | Practice challenges.

Related Power of AI pages

Keep reading with Start here, everyday uses, the tool guide, the writing workshop, and sources and standards.