How to Write Instructions an AI Agent Can Follow

Ever asked an AI agent to do something and gotten back... not quite what you meant? Usually it's not the agent. It's the instructions. The good news is that writing good ones is a skill you can pick up in an afternoon.

Think of it like training a new hire who's smart, fast, and has never met you. They'll do exactly what you say, so say it clearly.

Here's a simple recipe, step by step: First, give it a role. One line on who it is, like "You're my customer follow-up assistant." Amazon's Bedrock example does exactly this: "you are an inventory management agent." Next, state the goal. What does done look like? One or two sentences. Then, give it the context. Who it's for, what matters, and anything it needs to know about your business. After that, list the steps, in order, if the task has them. Now, set the rules. What it should never do, like inventing prices, or sending without your OK. And last, show an example. One good example of the output beats a paragraph of description.

Here's a template you can copy:

• Role: You are my [job] assistant.
• Goal: [what finished looks like].
• Context: [who it's for, what matters].
• Steps: [1, 2, 3].
• Rules: Only use [sources]. If you're not sure, write "unknown." Don't send, buy, or delete anything without my approval.
• Format: [list, table, short email].
• Example: [paste one you like].

The most common mistakes:

• Too vague: "Help with marketing." Name the task and the finish line.
• Too many jobs at once: give each agent one clear job.
• No permission to say "I don't know": tell it that "unknown" is a fine answer.
• No example: if you have one, include it.
• No approval rule: say plainly what needs your OK.

The big platforms work this way, too. Salesforce says you write natural-language instructions for each Agentforce subagent, and Voiceflow's Agent Studio is where you shape an agent's persona, tone, and instructions. Grok Bot users can see our guide to briefing a Bot for the same ideas.

And here's a tip Microsoft puts right in its docs: if a simple fixed step can handle a task, use that instead of an AI agent. Save the agent for the parts that need judgment.

Start with the template, run it a few times, and tweak one line at a time. You'll be surprised how fast it gets good.