Writing an AI agent's mission: the five-point method
An agent works only as well as its mission is clear. This course gives you a five-point method to write it: the purpose, the inputs, the outputs, the limits, and the tone. You will see failed missions rewritten, and you will understand why a good agent must question you before acting.
- A clear mission states the purpose, the inputs, the outputs, the limits, and the tone.
- Limits matter as much as purpose: write what the agent must never do.
- A well-described output can be verified at a glance.
- A good agent questions you before acting, and that's a good sign.
When an agent disappoints, the reflex is to blame AI. Most often, it's the mission that was unclear. A new employee told to "take care of customers" would do anything too. This course does not talk about magic formulas. It offers a simple framework to say what you expect, as you would for a serious apprentice.
IMission or prompt: what's the difference?
The written description of what an agent must do, starting from what, to produce what, and within what limits.
A prompt is a one-time instruction given to an assistant like ChatGPT for an immediate response. A mission serves each time the agent runs, for weeks. It must therefore hold up without you being there to clarify.
This is the essential difference. In a conversation, you correct as you go. An agent that works every Monday only has its mission and its documents. Everything you have not written, it will have to guess, and an agent that guesses ends up making mistakes.
IIThe five points of a good mission
- 1The goal
One sentence with an action verb and an expected result. "Prepare the delivery schedule for the following week every Friday", not "help me with deliveries".
- 2Inputs
What the agent reads to work: which documents, which emails, which list. Also say which version is authoritative if two files contradict each other.
- 3The outputs
What it produces, in what form, for whom. A ten-line email, a quote in the usual format, a summary table. The more precise it is, the more verifiable it is.
- 4The limits
What it never does and the cases where it stops to alert you. The amounts, the subjects, the clients, the days to avoid.
- 5The tone
Formal or informal tone, length, phrases to avoid, signature. Attach one or two texts you have written that sound like you.
The most neglected point is that of the limits. We think about what the agent must do, rarely about what it must not do. Yet that is where annoying errors hide: a discount given, a customer followed up during a dispute, a message sent on a Sunday evening.
IIIThree failed missions, rewritten
Here are three typical missions, as they are often written in small businesses, then their corrected version. Note that the good version is not much longer: it is simply more precise.
IVDescribing the output so you can verify it
A well-described output saves you time during validation. If you know what the result should look like, you immediately see what's wrong. Specify the length, structure, and required elements. For a follow-up email, for example: invoice number, amount, due date, payment method, polite closing. If an element is missing, you spot it in five seconds instead of rereading the text three times.
Check what applies to you: the list remembers your answers on this device.
VWhy an agent must ask for what it needs
Even a good mission leaves gaps. You write "follow up on overdue invoices", but what about a customer who has paid half? An agent who assumes will choose for you. An agent who asks forces you to make a decision once and for all. A real estate agency that entrusts the response to visit requests will think about schedules and available properties. The agent, on the other hand, will ask what to say to someone who wants to visit a property already under contract. This case happens, and no one had written it down. This is why an ABYS agent demands everything it needs before working: its questions are not a nuisance, these are the cases you had not anticipated.
VIEvolving the mission without breaking everything
A mission is not set in stone. Change one thing at a time, then look at the next passage. If you change the tone, deadlines and customer list all at once, you won't know what produced the change. On ABYS, a dry run allows you to check the effect of a change without anything being sent. I advise you to note, in one line, each change and its date: this is your memory when a result surprises you.
A mechanic writes: "Every Monday, from the attached list, prepare a ten-line email per overdue invoice, using formal address." Which point of the method is missing?
?Questions we're asked
How to write a good prompt for an AI?
State the goal in one sentence, give the context and useful documents, describe the expected result, specify what must not be done and the desired tone. Give an example of what you like. For an agent working alone for weeks, these five points become essential.
How long should an AI agent's mission be?
Long enough to cover the goal, inputs, outputs, limits and tone, often ten to twenty lines. A mission that is too short leaves the agent guessing. A mission that is too long mixes the essential and the secondary: reference details belong in attached documents.
Do I need to write the mission in English for it to work better?
No. Current agents understand French very well. Write in the language of your customers and your documents: you will review the mission more easily, and the tone of the messages produced will be more natural.
Why isn't my AI agent doing what I want?
Most often, the mission leaves room for interpretation: a vague goal, a missing document, an unwritten limit. Review it with the five points, compare with the result obtained, then correct only one element at a time and run a dry run again.
Do you have your mission in five points? Describe it in your own words: the agent will ask you the missing questions before proposing its plan.
Create My AgentRead more next
Create your first AI agent in ten minutes, without installing anything
By the end of this tutorial, you will know how to create an AI agent that handles a specific task from your week. You will know what to…
Read the lessonWhat documents to give an AI agent, and which to keep
An AI agent works from what you give it. This guide tells you which documents really help it, in what form…
Read the lessonThe eight mistakes that make an AI agent fail, and how to fix them
When an AI agent disappoints, the cause is rarely technical. It's almost always a vague instruction, a forgotten document or a…
Read the lesson