How to Prompt AI Agents: Write a Brief, Not a Question

Prompting an AI agent is closer to briefing a new colleague than asking a chatbot a question. A good agent prompt states the goal, explains why it matters, defines what "done" looks like, sets limits on what the agent may do, names the tools or sources to use, and says when to stop and report back. Chat prompts can get away with one line; an agent that will take ten steps on its own needs those six things, or it will fill the gaps with guesses.

For one-off chat prompts, our tutorial on writing better AI prompts covers the role-context-task-format basics. This guide is about the extra work an agent needs.

Chat prompt vs agent brief

What the model makers advise

Both major vendors publish prompting guidance, and the agent-relevant points line up (checked 16 October 2026):

• Anthropic's "golden rule": show your prompt to a colleague with minimal context and ask them to follow it — "if they'd be confused, Claude will be too." Anthropic also recommends giving the context or motivation behind instructions, and notes its latest models follow instructions precisely: ask "can you suggest some changes" and you may get suggestions rather than changes.
• Anthropic on long tasks: make steady progress on a few things at a time, keep track of state in structured formats, and give the agent ways to verify its own work.
• OpenAI's guidance for agentic workflows: tell the agent to plan and break the request into sub-tasks, keep going until it is fully resolved, explain why before calling a tool, and use a to-do list or rubric so steps are not skipped. OpenAI also suggests organising instructions into identity, instructions, examples and context.

The six-part agent brief

1. Goal. One sentence on the outcome, not the method. "Produce a shortlist of three accounting tools for a five-person firm."

2. Context and why. Who it is for and what matters to them. "The firm works in two currencies and the owner is not technical, so ease of setup matters more than features."

3. Done-criteria. How you will judge the result. "Each option has a current price from the vendor's own page, a source link and one sentence on fit."

4. Limits. What the agent must not do. "Do not sign up for trials or enter any details into forms. Use only vendor sites and official documentation."

5. Tools and sources. Name them explicitly — agents follow instructions literally. "Use web search and the vendors' pricing pages. Put results in the attached spreadsheet."

6. Checkpoints. When to stop and show you. "After the long-list of ten, stop and wait for my go-ahead before narrowing to three."

A worked example: before and after

Before: "Find me the best accounting software."

What usually comes back is a generic list with no prices, no sources and no idea of your situation.

After:

"Goal: shortlist three accounting tools for a five-person consultancy. Context: we invoice in GBP and EUR; the owner is not technical, so setup time matters more than advanced features. Done when: each option has its current price from the vendor's pricing page (with the date), a link to that page, multi-currency support confirmed from the vendor's documentation, and one sentence on fit. Limits: read-only — no sign-ups or form entries; vendor sites and official docs only. Tools: web search; write results into a table with those columns. Checkpoint: show me a long-list of up to ten first and wait before narrowing it down. If a fact cannot be confirmed on a vendor page, write NOT FOUND rather than guessing."

The last sentence matters most: it gives the agent permission to admit a gap instead of inventing a number.

Keeping long tasks on track

• Ask for a running to-do list and have the agent tick items off as it goes, so you can see where it is.
• Have it keep notes outside the conversation — a file or table — for anything it must remember across steps.
• Ask it to state why before each tool call while you are still building trust.
• Plan the brief before you write it. Mapping goal, limits and checkpoints in a tool such as MindManager helps you spot what is missing.
• Give it material, not just instructions. For document-heavy research, tools such as PopAI accept PDFs, Word files and web pages as inputs.

Common mistakes

• Asking for suggestions when you want action — or the reverse. Say which you want.
• No stopping point. An agent without done-criteria either stops too early or wanders.
• Hidden constraints. If budget, region or tone matters, say so; the agent cannot read your mind.
• Too many jobs in one brief. Research, write and publish are three briefs with checkpoints between them.
• No permission to say "I don't know". Without it, gaps get filled with plausible guesses.

Frequently asked questions

How is prompting an AI agent different from prompting a chatbot?

A chatbot gives one answer; an agent takes many steps and may use tools. So an agent prompt needs a goal, done-criteria, limits on what it may do, the tools to use and checkpoints — not just a question and a format.

What should an AI agent prompt include?

Six parts work well: the goal, context and why it matters, how you will judge "done", what the agent must not do, which tools or sources to use, and when to stop and report back.

How do I stop an AI agent from making things up?

Require sources for every fact, tell it to write "not found" when it cannot confirm something, and check a sample of its sources yourself. Asking it to verify across more than one source also helps.

How long should an agent prompt be?

As long as the job needs. A six-line brief that covers goal, context, done-criteria, limits, tools and checkpoints usually beats both a one-liner and a page of step-by-step instructions.

Related

Put briefs to work in how to build an AI agent workflow, keep access tight with are AI agents safe?, and find the right tool in our best AI agents overview or with the Agent Finder. All tutorials live in the agent tutorials category.

Sources checked 16 October 2026: Anthropic, "Prompting best practices" (platform.claude.com); OpenAI, "Prompt engineering" guide (developers.openai.com). Editorial content.