Intention: the mission and the why
What it is
Intention is the first thing an agent needs to receive: what should be achieved and why it matters. The task tells what to do; the why tells what to protect when a case shows up that you didn’t anticipate. In the Agent Sheet, this is the field “What do you want to take out of your hands? Why does that matter?”. An agent with the right task and no why executes well the wrong thing.
Why learn
No instruction covers all cases. When a doubt arises, the agent needs a criterion to choose, and the why is that criterion. In a clinic, “confirming consults” without a reason turns into sending a message and done; with “30% of patients don’t show up and the schedule stays empty,” it understands that the important thing is knowing who won’t come in time to fit someone else in.
Key concepts
Goal: give the agent the mission and the reason, so they can decide well in unexpected cases.
## Your mission
<task, e.g.: Confirm the consultations from the next day via WhatsApp>.
This matters because: <real issue, e.g.: 30% of patients don’t show up without notice and the calendar is left empty>.
When you’re in doubt, prioritize this.
Before you start, confirm in 3 lines what you understood about your mission.
How to check: read the 3-line confirmation. If they don’t repeat the reason in their own words, the mission is still too vague.
✓ Do this
✓ Write the why as a painful problem, with a number if you have one: “I lose 2 hours a day”.
✗ Avoid this mistake
✗ Use generic whys like “improve customer service,” which don’t help you decide anything.
Practice before you reveal
Choose a task you want to delegate and write the mission in two lines: the task and the why.
View commented answer
Store example: “Reply to customers about order status. This matters because I lose 2 hours a day with the same question and the customer keeps waiting.” The why shows the response has to be fast and correct, not just a nice-sounding text.
One task per agent
What it is
Scope is the size of the work you deliver: where the task starts and where it ends. “Taking care of my business” isn’t a task—it’s a whole job with dozens of different decisions. A good agent gets only one task, with clear input and output, like “check the month’s documents and list what’s missing.”
Why learn
When an agent does five things and gets some wrong, you don’t know which part failed. With one task per agent, each mistake has an address, each test is simple, and each autonomy level can be different: what only answers questions can have more freedom than what handles money.
Key concepts
Goal: split a big request into tasks that you can delegate one at a time.
Big request: <e.g.: handle store customer service>
Delegate-able tasks (one per agent):
1. <e.g.: respond to order status and delivery time>
2. <e.g.: answer exchange questions within 7 days>
3. <e.g.: compile the daily summary of the most common questions>
Stay with me (not delegating yet):
- <e.g.: refunds, discounts, complaints>
How to check: each task on the list fits into a single sentence with a verb, has an output you can see, and doesn’t depend on you approving money.
From concept to action
- Take care of the business: identify the initial condition.
- Customer service: apply the described decision.
- Order status: check the effect in the example.
- Response sent: record the output evidence.
✓ Do this
✓ Start with the most repeated, lowest-risk task, and only then create the second agent.
✗ Avoid this mistake
✗ Set up an “all-purpose agent” with a giant instruction that nobody can test.
Practice before you reveal
Take one area of your work and split it into three delegable tasks and one that stays with you.
View commented answer
Accountant office example: check documents received; list what’s missing by customer; set up the reminder for pending items. Stays with you: talk to the customer about the fine. The one that stays is always the one with the highest risk or the one that requires judgment.
When priorities clash
What it is
Every real task has priorities that clash: speed versus accuracy, friendliness versus firmness, doing everything versus doing it right. The tie-break rule is a sentence that says which priority wins when the two can’t both fit together. It’s written in the instruction, not in your head.
Why learn
Without a tie-break, the agent chooses on its own, and the choice changes from day to day. A support agent who answers quickly with an invented deadline pleases today and creates complaints tomorrow. Writing the order of priorities makes behavior predictable and easier to correct.
Key concepts
Goal: tell the agent which priority wins when two conflict.
## When priorities conflict
- <priority A, e.g.: Precision> comes before <priority B, e.g.: speed>.
In practice: <e.g.: I’d rather respond within 1 hour with the correct deadline than within 1 minute with a guessed deadline>.
- <priority C, e.g.: follow the house rules> comes before <priority D, e.g.: please the customer>.
- If no rule resolves it: don’t decide alone—call me and tell me what the conflict was.
How to check: invent a case where both priorities conflict and check whether the agent picks the one you wrote first.
✓ Do this
✓ Write the tiebreaker using a concrete example from your business right below it.
✗ Avoid this mistake
✗ List ten priorities that are “all important” without saying which one wins.
Practice before you reveal
Write a tiebreak rule for your agent in the format “X comes before Y” and provide an example.
View commented answer
Professor example: “Fairness to the student comes before finishing fast: if the writing is illegible or off-topic, mark review and don’t give a grade.” Notice that the rule says what to do, not only what to value.
Context: what you’d explain on the first day
What it is
Context is everything the agent needs to know about your environment to decide well: house rules, tone, schedules, what you never say to the customer. The best way to gather this is to imagine the first day of a new employee—what, in a company, is called onboarding, the integration of whoever is arriving. What you’d explain at the door, the agent also needs to hear.
Why learn
The agent doesn’t know your clinic, your store, or your team. Without the house rules, it fills the gap with the internet’s default: it messages on Saturday, calls the patient for you, promises what it can’t. A short block of rules prevents most of those silly mistakes.
Key concepts
Goal: give the agent what you would explain to a new employee on day one.
## House rules
- <e.g.: address people as sir or ma’am>
- <e.g.: messages only between 9:00 and 17:00>
- <e.g.: Saturday has no appointments>
- <e.g.: don’t discuss prices; that’s with me>
- Tone: <e.g.: polite, short, no slang>.
Ready sheet to compare: https://inematds.github.io/7pa/en/?preset=clinica
How to check: ask the agent to write a message on a Saturday at 8:00 PM and see if they refuse based on the time.
✓ Do this
✓ Write the rules as short sentences that a new employee would understand without having to ask.
✗ Avoid this mistake
✗ Paste the entire company manual and expect the agent to find what matters.
Practice before you reveal
List five things you would tell someone on the first day doing the task you want to delegate.
View commented answer
A good list mixes rule, tone, and limits. Real estate example: don’t share the exact address before the visit, respond within up to 1 hour, upbeat and to-the-point tone, don’t negotiate the price, visits only during business hours.
The smallest context that makes a good decision
What it is
More context isn’t better context. Context engineering is choosing what goes into the agent’s instruction and cutting the rest. The cutting question is simple: does this information change any of its decisions? If it doesn’t, it goes out.
Why learn
Too-long instructions hide important rules in the middle of text that doesn’t matter, and the agent starts giving the same weight to the company’s history and the rule about not talking about price. A lean context is easier to read, test, and fix when something goes wrong.
Key concepts
Goal: trim the agent instruction, keeping only what changes decisions.
For each line of your instruction, answer:
- Line: <e.g.: The clinic was founded in 1998>
Does it change any agent decision? <no> → CUT
- Line: <e.g.: Saturday has no appointments>
Does it change any agent decision? <yes, doesn’t confirm an appointment on Saturday> → KEEP
How to check: run the same request with the old instruction and with the trimmed one. If the answer is still correct, the cut was a good one.
From concept to action
- Bring everything together: identify the initial condition.
- Does it change the decision?: apply the described decision.
- Cut the rest: check the effect in the example.
- Test without: record the output evidence.
✓ Do this
✓ Cut one line at a time and test, to see which cut made the difference.
✗ Avoid this mistake
✗ Keep the text “just to be safe,” as an institutional mission and company history.
Practice before you reveal
Take an instruction you already use and, line by line, mark what changes a decision and what doesn’t.
View commented answer
Usually a third of the instruction comes out. What remains is rules, limits, and the source of the data; presentations, praise, and explanations that the agent doesn’t use get removed. If, after cutting something, it starts making mistakes, that line goes back.
Examples are worth more than adjectives
What it is
Adjectives like “cordial,” “professional,” and “objective” mean different things to each person. An example shows exactly what you want. The simplest technique is to put in the instruction a good response and a bad response for the same case, explaining why one is good and the other isn’t.
Why learn
An agent copies patterns much better than it interprets adjectives. A model answer teaches tone, length, the order of information, and what not to say—all at once. And the bad answer shows the mistake you fear most, which usually helps avoid repeating it.
Key concepts
Goal: teach the tone with a good and a bad example, instead of adjectives.
## Response example
Case: <e.g.: the customer asks where order 1234 is>
GOOD (do this):
“<e.g.: Hi, Marina! Your order 1234 shipped yesterday and delivery is scheduled for Thursday. If you need anything, just reach out.>”
BAD (don’t do this):
“<e.g.: Dear customer, we inform you that your order is in transit and should arrive soon.>”
Why: the good one uses the person’s name, the order number, and a system date; the bad one is cold and promises “soon” without any data.
How to check: ask three similar questions and confirm the answers follow the good model.
✓ Do this
✓ Use a real case that happens every week, with example names and numbers.
✗ Avoid this mistake
✗ Write “be friendly, but professional and to the point” and expect that to be enough.
Practice before you reveal
Write a pair of examples—one good answer and one bad one—for the most common case of your task.
View commented answer
The pair works when the difference is visible in one line. Salon example: the good one says “cut + blow-dry for R$ 90, book via the link”; the bad one says “can’t-miss prices, come check it out.” The bad one is generic and doesn’t include the data the customer wants.
Check your understanding
Your customer support agent is torn between responding quickly with an approximate timeframe or waiting to confirm in the system. What solves this?
What you take from this module
Write the mission, the tie-break rule, and the “House rules” block of an agent—ready to paste into ChatGPT or into Claude.
- Intention: the mission and the reason.
- One task per agent.
- When priorities clash.
- Context: what you would explain on day one.
- The smallest context that makes you decide well.
- Examples matter more than adjectives.
Next action: apply what you learned to your agent sheet and note what still needs review.