When an AI agent disappoints, the technology is rarely to blame. The problem is almost always an instruction that's too vague: “follow up with my clients,” “make me a report,” “sort my emails.” Those sentences are clear to you because you know the context. They aren't clear to the agent.
Writing a good brief takes no technical skills. It's the same exercise as explaining a task to a capable new team member who doesn't yet know your clients, your tools or your habits. This tutorial gives you a template, examples and a way to check your work.
Why the brief matters so much#
An AI agent makes decisions on every run: which data to read, which cases to handle, what tone to use, what to do when something unexpected comes up. Every gray area in your instruction is a decision it will make alone, based on its own interpretation. A good brief shrinks those gray areas down to the cases where you're fine with it improvising.
On Future Agent, you're not left staring at a blank page: the AI architect reads your description and asks about whatever is missing. But the more precise your first description, the fewer back-and-forths, and the more faithfully the agent does what you want.
The seven-part template#
An effective brief answers seven questions. You don't need headings or formatting: a few sentences per part is enough.
1. The goal#
Why does this task exist? “Never leave a quote without a follow-up” is a goal. “Send emails” isn't. The goal helps the agent make the call on cases you didn't anticipate.
2. The trigger#
When should the agent work? At a set time (“every Monday at 8 a.m.”), on an event (“with every new form submission”), after another agent, or on demand. Specify the time zone if your team is spread out.
3. The sources#
Where is the information? Be specific: file name, tab, mailbox folder, label, the column that holds the status. “In my Drive” isn't enough; “in the Quote tracking 2026 file, In progress tab” is.
4. The rules#
What to do, and above all what not to do? This is the most important part. List the selection criteria, exceptions and edge cases:
- which items to handle (status, date, amount, client type);
- which to ignore (disputed accounts, public sector, amounts above a threshold);
- what to do when in doubt (usually: do nothing and let you know).
5. The expected output#
What should the agent produce, and where? An email sent, a draft, a row added, a Slack message, a PDF or an Excel file. Describe the format: length, sections, columns, subject line.
6. Tone and style#
Formal or casual, buttoned-up or warm, sign-off, phrases to avoid. The simplest approach is to paste one or two messages you've written and are happy with.
7. Limits and approvals#
What the agent must never do without your approval: email a client, delete a file, change accounting data, spend money. Also say what kind of summary you want to receive.
A template to copy#
Fill in the brackets and paste the text into the chat:
- Goal: [why this task exists]. - When: [time, day, triggering event]. - Where to find the information: [tool, file, tab, folder, column]. - What to do: [main steps]. - What not to do: [exceptions, cases to ignore]. - When in doubt: [do nothing and tell me / other]. - Output: [what to produce, where and in what format]. - Tone: [formality, style, sample message]. - Approval: [what must be sent to me before it goes out]. - Summary: [what I want to receive, and how often].
Before-and-after examples#
Invoice reminders#
Before: “Chase my clients who haven't paid.”
After:
Goal: get paid on time without damaging client relationships. Every Tuesday at 9 a.m., check invoices with the status “Sent” in my invoicing software. For those more than 7 days overdue, draft a first courteous reminder that states the invoice number, amount and due date, with the payment link. For those more than 21 days overdue that have already had one reminder, draft a second, firmer but still polite reminder. Never write to clients marked “Dispute” or to government bodies. If you're unsure about an account, do nothing and flag it. Professional, formal tone, signed “Accounts team.” Send me the reminders as drafts on Tuesday morning; I'll send them myself during the first month.
The second version contains the goal, the trigger, the source, two graduated rules, exclusions, what to do when in doubt, a tone and an approval step. The agent has almost nothing left to guess. Related template: Unpaid invoice reminders.
The Monday report#
Before: “Make me a report every week.”
After:
Every Monday at 7:30 a.m., prepare my weekly update. Take last week's invoiced revenue from my invoicing software, new opportunities and deals won or lost from the CRM, and the bank balance from the Cash tab of my spreadsheet. Compare each figure with the previous week and with the same week last month. Put the three items that need my attention at the top (significant variance, stalled deal, late payment). One page max, as a PDF, emailed to me and my business partner. No commentary on anything that's stable.
Related template: Weekly management report.
Email triage#
Before: “Sort my emails.”
After:
Every morning at 8 a.m., go through the emails received since yesterday in my main inbox. Sort them under three labels: “To do” (a request or question addressed to me), “To read” (useful information, no action needed), “To archive” (notifications, ads, newsletters). Never delete anything. Send me a summary with the three most urgent emails and one sentence for each. Messages from my accountant and my bank always go to “To do.”
Related template: Inbox triage.
The questions the architect will ask#
Even with a good brief, the Future Agent architect will ask a few questions. Anticipating the most common ones saves time:
| Question | Why it matters | Sample answer |
|---|---|---|
| Which tool holds the data? | To request the right connection | “Google Sheets, Client tracking file” |
| How often? | To set the trigger | “Monday to Friday at 9 a.m.” |
| Send or draft? | To set the level of autonomy | “Drafts for a month, then send” |
| Who receives the output? | For delivery | “Me and the office manager” |
| What format? | For the deliverable | “A one-page PDF and the detailed Excel file” |
| What to do when in doubt? | For safety | “Do nothing and ping me on Slack” |
The architect then shows you a sample of the result. That's the moment to fix the tone, the order of information or a misunderstood rule, before anything goes live.
The most common mistakes#
Assuming the agent knows the context#
“Follow up with the usual clients”: which ones are usual? By what criterion? If the criterion isn't written down anywhere, add a column to your file or list them in the brief.
Forgetting what not to do#
Exclusions prevent nasty surprises. Think of disputed accounts, personal contacts, closure periods, unusual amounts.
Mixing several tasks#
“Sort my emails, follow up on my quotes and prepare my report”: that's three agents. Kept separate, they're easier to configure, monitor and fix. They can then pass information to each other, as explained in our article on multi-agent workflows.
Being vague about format#
“A summary” could be three lines or three pages. Specify a length, sections and an order.
Not providing a tone example#
One message you've written is worth more than a long description of your style. The agent then remembers your preferences thanks to its memory: see how AI agents remember your preferences.
Attaching context documents#
Some information doesn't fit in a brief: your price list, your terms and conditions, a job description, your product catalog. On Future Agent, you can add context files to an agent (PDF, Word, Excel, CSV, text) by dropping them on its card in the dashboard or through the “Context” button. The agent consults them when useful, without you having to copy them into the instruction.
The pre-launch checklist#
Before approving, reread your brief with these questions:
- Would a team member who started this morning know what to do with this text alone?
- Is the trigger precise (day, time, event)?
- Are the sources named exactly?
- Are the exclusions listed?
- Is the when-in-doubt behavior defined?
- Is the output format described?
- Is it clear what needs human approval?
Seven yeses and you're ready to launch. If not, fill in the missing part: it takes two minutes and will save you far more.
Evolving the brief#
A brief isn't set in stone. After the first runs, you'll notice cases you didn't anticipate. Describe the fix in the agent's chat, in one sentence: “From now on, don't follow up on quotes over €10,000, I'll handle those myself.” The agent applies the new rule from the next run.
If you don't know where to start, pick a template from the library: its brief is already written, and the architect only asks about what's specific to you. To see every available setting, visit the Features page.
In short#
- A good brief answers seven questions: goal, trigger, sources, rules, output, tone, limits.
- Exclusions and when-in-doubt behavior matter as much as the steps.
- A sample message beats a description of your style.
- One task, one agent.
- The brief improves with the first runs: fix it one sentence at a time, as you go.