AI Agent Instructions: How to Write Them (With Template)

Short answer Good AI agent instructions read like a brief for a capable new colleague: say what the goal is and why, give the context the agent can’t guess, set limits (budget, sources, dates, what not to touch), describe what “done” looks like, and say which steps need your approval before it acts. OpenAI, Anthropic and Google all publish guidance along these lines. Vague requests such as “handle my inbox” are the most common cause of disappointing results.

An AI agent is only as useful as the instructions it’s given. Ask a chatbot a loose question and the worst outcome is a loose answer. Give an agent a loose instruction, and it may spend twenty minutes clicking through websites, booking the wrong thing, or stopping halfway to ask what you meant.

The good news is that writing clear instructions is a learnable skill, and the companies building these agents have published fairly consistent advice about it. This guide pulls together that official guidance from OpenAI, Anthropic and Google, and turns it into a practical method with templates you can copy. It’s based on publicly documented guidance as of September 2026, not on our own testing.

If you’re new to agents, our beginner’s guide to AI agents explains what they are and how they differ from chatbots.

Why do AI agent instructions matter more than chatbot prompts?

A chatbot answers. An agent, on the other hand, acts. It may browse the web, fill in forms, read your email, update a spreadsheet or run for a long time without checking in. That changes the stakes of an ambiguous instruction in three ways:

Errors compound. A wrong assumption in step one carries through every later step.
Actions can be hard to undo. A sent email or a completed booking can’t simply be regenerated.
You’re often not watching. Agents are designed to work while you do something else, so they can’t read your reaction and adjust.

OpenAI’s help page for ChatGPT agent makes the point directly, advising users to avoid vague requests like “Check my email and handle everything.” Anthropic’s prompting guide offers a helpful framing: treat the AI as a “brilliant but new employee” who lacks context on your norms and workflows. The more precisely you explain what you want, the better the result.

The six parts of good AI agent instructions

Across the official guides, the same ingredients keep appearing. So we’ve grouped them into six parts. Not every task needs all six, but anything that involves the agent taking actions on your behalf usually benefits from each one.

Part	What it answers	Example
1. Goal	What are you trying to achieve?	“Find a hotel for two nights in Lisbon”
2. Why and context	What’s it for, and what does the agent need to know?	“It’s for a work conference at the FIL venue; I don’t have a car”
3. Constraints	What are the limits?	“Under €180 a night, free cancellation, walkable to the venue”
4. Steps or approach	Is there an order or method to follow?	“Check at least three booking sites and compare”
5. Output	What should the result look like?	“A table of three options with price, distance and cancellation terms”
6. Checkpoints	Where should it stop and ask?	“Don’t book anything; show me the options first”

1. Start with the goal, stated plainly

Google’s Gemini API documentation says instructions can range from a simple question to a step-by-step task. Either way, what matters is being explicit. Anthropic’s guide adds a point that’s particularly relevant to agents: if you want action, ask for action. “Change this” gets different results from “Can you suggest changes?” With an agent, decide whether you want it to research and report or to actually do the thing, and say which.

2. Explain why, and give the context it can’t guess

This is the step most people skip. Anthropic’s guide recommends explaining the reason behind an instruction rather than just stating a rule, because the AI can then apply the reason to situations you didn’t anticipate. Its example: rather than “never use ellipses,” explain that the response will be read aloud by a text-to-speech engine that can’t pronounce them.

The same applies to everyday tasks. “Find a restaurant” is a guess waiting to happen. “Find a restaurant for my parents’ 40th anniversary; my dad uses a wheelchair and my mum is vegetarian” gives the agent what it needs to make sensible choices on its own.

Google’s Workspace guidance lists “provide context” among its core tips, and its developer documentation suggests including reference material where relevant, so the answer is tailored rather than generic.

3. Set constraints: budget, dates, sources, boundaries

Constraints are what stop an agent drifting. Google’s documentation gives the example of specifying a precise length for a summary, but the principle scales up: price ceilings, date ranges, preferred or banned websites, brands to avoid, locations, deadlines.

For agents with access to your accounts, add boundaries on scope. OpenAI’s agent guidance recommends enabling only the apps a task needs. You can reinforce that in the instruction itself: “Only look in my Work calendar” or “Don’t open any emails older than a month.”

A tip from Anthropic’s guide: say what to do rather than only what not to do. “Write in flowing paragraphs” works better than “don’t use bullet points.” Positive instructions give the agent a target instead of just a fence.

4. Break big jobs into steps

Both Google and Anthropic recommend breaking complex requests into smaller pieces, either as numbered steps in one instruction or as a sequence of separate prompts. Anthropic suggests numbered lists when the order matters.

For an agent, steps also make the work easier to check. If step two produces something odd, you can stop there rather than discovering the problem at the end.

5. Describe the finished product

Tell the agent what “done” looks like. Google’s documentation notes that you can ask for a table, a bulleted list, a short pitch, keywords, a sentence or a paragraph. For agents, add the practical details: should it include links? Prices with the date checked? A short summary at the top? A note of anything it couldn’t find?

If you have an example of what you want, include it. Google calls these “few-shot” examples and says they help control formatting, phrasing and scope. Anthropic recommends three to five varied examples for more complex tasks.

6. Say where it must stop and ask

This is the most important part for anything involving money, messages or deletion. Anthropic’s guidance for agentic systems suggests encouraging the AI to take reversible actions freely, but to ask before actions that are hard to undo, affect other people or could be destructive, such as deleting files or sending messages.

Most consumer agents already pause for some high-impact actions. OpenAI says ChatGPT agent asks for confirmation before consequential steps. But built-in checkpoints vary between products, and you know better than the agent where your own lines are. Stating them costs one sentence: “Don’t buy, send or delete anything; stop and show me first.”

AI agent instructions before and after: three worked examples

Shopping

Before: Find me a good vacuum cleaner.

After: I’m looking for a cordless stick vacuum for a two-bedroom flat with hard floors and one long-haired cat. Budget up to £250. Look at manufacturer pages and at least two retailers. Give me a table of three options with price, battery life as claimed by the maker, weight and where you found the price, with the date you checked. Note any common complaints you find in professional reviews. Don’t add anything to a basket.

The second version sets the use case (hard floors, pet hair), the budget, the sources, the output and a clear stopping point. It also asks for the date on prices, which matters because prices change. For more on shopping agents, see can AI agents find better prices?

Email and calendar

Before: Sort out my calendar for next week.

After: Look at my Work calendar for Monday 5 to Friday 9 October. List any double bookings and any day with more than five hours of meetings. For each clash, suggest which meeting to move, based on this rule: client meetings take priority over internal ones. Draft, but don’t send, a short polite message for each meeting that would need moving. Don’t change any calendar events.

This version limits scope to one calendar and one week, gives the agent a rule for decisions, and separates drafting from sending.

Research

Before: Tell me about new laws on AI.

After: I run a small online shop in Ireland and use an AI chatbot for customer questions. Find official EU or Irish government sources explaining whether I need to tell customers they’re talking to an AI, and from what date. Quote the relevant line and link to the source. If sources disagree or the answer isn’t clear, say so instead of guessing. Keep it under 300 words.

Asking the agent to say when something is unclear, rather than filling the gap, is one of the most useful lines you can add to any research task.

A copyable template for AI agent instructions

Paste this into your agent and fill in the brackets:

Goal: [what you want done]. Context: [why, who it’s for, anything the agent can’t guess]. Constraints: [budget, dates, locations, sources to use or avoid, accounts it may use]. Steps: [any order to follow]. Output: [format, length, what to include, e.g. links and dates checked]. Checkpoints: before you [buy / send / book / delete / change anything], stop and show me. If something is unclear or you can’t find it, tell me rather than guessing.

It’s longer than most people are used to typing, and that’s the point. Google’s Workspace guidance encourages writing to Gemini in natural language, as you would to a colleague. A few extra sentences at the start tend to save time later.

How do you test and improve AI agent instructions?

OpenAI’s ChatGPT guidance describes prompting as iterative: try an instruction, look at the result, then refine the wording, add context or simplify. For agents, a few habits help:

Use Anthropic’s “colleague test.” Its guide suggests showing your instruction to someone with little context on the task. If they’d be confused, the AI probably will be too.
Run it once while watching. Before you schedule a task or leave an agent unattended, run it manually and see where it hesitates or goes wrong. This is especially relevant for scheduled AI tasks.
Save what works. When an instruction produces good results, keep it. Many assistants now let you save instructions for reuse, as projects, skills or plugins.
Change one thing at a time. If results are poor, adjust one element (the context, a constraint, the output format) so you can tell what made the difference.

What instructions can’t fix

Clear instructions improve results, but they don’t make an agent infallible. Agents can still misread websites, work from outdated information, or be misled by content they encounter. A particular risk is prompt injection, where text on a web page or in an email tries to give the agent its own instructions. OpenAI’s agent documentation acknowledges that its protections reduce this risk but don’t eliminate it.

So treat your instructions as one layer of protection, not the only one. Keep sensitive actions behind your own approval, connect only the accounts a task needs, and check the output, especially facts, prices and anything the agent says it has done.

Common mistakes in AI agent instructions

Assuming shared context. The agent doesn’t know your budget, your city or that your “work calendar” is the second one.
Bundling unrelated jobs. “Plan my trip, clear my inbox and update the budget” is three tasks. Run them separately.
Only saying what not to do. Give a target, not just a list of prohibitions.
No definition of done. Without one, agents tend to either stop early or keep going longer than you wanted.
No stopping point for irreversible actions. One sentence prevents most expensive mistakes.
Not asking for uncertainty. Tell it to flag what it couldn’t verify.

Key takeaways

Treat an AI agent like a capable new colleague: give it the goal, the reason and the context it can’t guess.
Set constraints (budget, dates, sources, accounts) and describe exactly what the finished result should look like.
Break big jobs into numbered steps so problems are caught early.
Always say where the agent must stop and ask, especially before buying, sending or deleting.
Test an instruction while watching, refine one thing at a time, and save versions that work.

AI agent instructions: FAQs

How long should instructions for an AI agent be?

Long enough to cover the goal, context, constraints, output and checkpoints. For a real task that is often a short paragraph or a few numbered steps, rather than a single line.

Should I tell an AI agent what not to do?

Set clear limits, especially on irreversible actions, but where possible phrase instructions positively. Anthropic’s guidance says telling the model what to do works better than only saying what to avoid.

Do the same instructions work in ChatGPT, Gemini and Claude?

Largely, yes. OpenAI, Google and Anthropic publish similar advice: be clear and specific, give context, specify the format and break complex tasks into steps. Results can still differ between products.

Can good instructions stop an agent making mistakes?

They reduce mistakes but can’t prevent them all. Keep approval steps for purchases, messages and deletions, and check important output yourself.

How to write good instructions for an AI agent (with templates)

Why do AI agent instructions matter more than chatbot prompts?