docs
Quick start

Best practices

Proven ways to get better work from your AI employees - a clear brief, the right first task, the right model, and a habit of reviewing.

Overview

The quality of an agent's work comes down to four things you control: the brief you give it, the task you start it on, the model it runs on, and how you review and correct it. This page covers each in turn.

Work through these in order. A vague brief causes most disappointing results, so fix that before reaching for a different model.

Write a brief that works

The brief lives in Employee settings → identity. The agent reads it before every task, in every channel, so it's the highest-leverage text in Opereo.

A good brief answers four questions:

  1. Where does it work? The company, the department, and what that department handles.
  2. What does it own? The two or three jobs it's responsible for, and what it should hand back to you.
  3. How should it sound? Tone, length, language, formatting habits.
  4. What must it never do? Guessing numbers, contacting customers without approval, touching money.

Here is a brief you can copy and adapt:

You are the customer support agent for Northwind Bikes, a bicycle shop in Leeds.

## Your job
- Answer customer questions about orders, repairs, and opening hours.
- Draft replies to support emails; I send them after review.
- Every Friday, summarise the week's common questions.

## Voice
Warm and short. Plain English. No exclamation marks.
Reply in the customer's language.

## Never
- Never promise a delivery date you can't see in the order system.
- Never offer a refund or discount. Flag it to me instead.
- Never guess at a price. Say you don't know and ask me.

Keep the brief about behaviour, not facts. Facts that change — price lists, policies, product sheets — belong in Files, attached to the conversation when they're needed.

Don't write the whole brief alone. Let the agent interview you about its own job and draft the brief from your answers — the method walks through it, and a Business Context Profile gives every agent the company-wide facts.

Start simple, then add

Start with a short brief and a single job. After a week of conversations you'll know exactly which rules are missing, because you'll have corrected the same thing twice. Add those rules then. A brief that tries to anticipate everything on day one is usually longer than it is useful.

Pick the right first task

Choose one job you repeat often and can check quickly. Good first tasks:

  • Drafting a weekly update or a customer email you'll read before sending.
  • Preparing research: "Find five suppliers of X and compare their minimum order."
  • Turning notes or a transcript into a clean summary.
  • Answering the same five questions your customers always ask.

Poor first tasks: anything irreversible (sending, paying, deleting) or anything you couldn't judge yourself in a minute.

Once the agent handles the first task well, give it the next one. Brief, review, expand.

Give it what it needs to read

Agents work better with the real documents than with your memory of them.

  • Upload price lists, policies, templates, and past examples to Files.
  • Attach the relevant file to the message when you ask. Agents only see a file when it's attached.
  • For tasks that touch your apps, connect the app instead of pasting data. A connected Gmail or Notion is always up to date; a pasted copy isn't.

Choose the right model

Every agent runs on a model you connect. Heavier models think harder and cost more of your plan or key; lighter ones are quicker and cheaper.

  • Lighter models for quick replies, summaries, formatting, and routine questions.
  • Heavier models for research, multi-step tasks, anything with numbers, and anything you'd hand to a senior colleague.

You can switch the model per agent from the model pill in the chat, and set Reasoning effort in the models settings for tasks that deserve more thinking. Details in Models.

Two agents in the same department can run on different models. A light model for the support agent and a heavy one for the research agent is a common setup.

Review, correct, repeat

An AI employee is still an employee. Read its work, especially in the first weeks.

  • Correct in the conversation. Tell it what was wrong and what you wanted. It remembers within that conversation.
  • Correct in the brief. If you find yourself correcting the same thing twice, it's a missing rule. Add one line to the identity.
  • Use the history. Every conversation is kept per agent. Open Chat history in the chat header to search past chats and see what it did.
  • Keep approvals human. Let the agent draft; you send. Let it propose; you decide. Give it write access to your apps only once it has earned it.

Use skills for repeatable know-how

If you catch yourself pasting the same instructions into every request — how to format an invoice, how to structure a report — turn them into a skill. Skills are reusable instruction packs the agent loads when they're relevant. You can install thousands from the hub or write your own in a minute. See Skills.

Automate only what's already working

An automation runs a prompt on a schedule while you're away. Automate a task only after you've run it by hand a few times and are happy with the result. Then have the automation deliver its output to a channel you actually read — WhatsApp, or your team's Slack.

Next steps

On this page