Rewire Agents

Use the BRIEF Framework to build your AI agents

Five questions you already answer to get a req approved and a new hire trained. Answer them for an agent and you have written the agent. The same document is a Claude project, a Custom GPT, a Gem, a Copilot agent, or an AGENTS.md file.

The five questions

Background
What is the background of the job? The context a new hire needs on day one: the team, the customers, the documents, how things are done here.
Role
What is the specific role, the task? One job, one agent. "You sort the shared inbox into three buckets." Not "you handle HR."
Instructions
What are the instructions to do the job well? The rules, the order of operations, and what needs a person before it goes anywhere.
Expectations
What does good look like? Three examples of good, three examples of bad. The field almost everyone skips, and the one that matters most.
Feedback
How do you make sure it gets better? Show it the final version and tell it why. Ask it what it assumed and what confused it. Update the BRIEF as you both learn. Put the review on the calendar. It can't learn from what it never sees.
If you were hiring someone tomorrow, you would tell them what it is about, what they do, how they do it, what good looks like, and how you are going to make sure they get better. That is an agent script. That's it.

One, filled in: the inbox sorter

Background. A shared HR inbox, about forty messages a day, three kinds: pay, policy, benefits. Two people currently read all of it.
Role. Read each message and say which of the three buckets it belongs in. Nothing else. Do not answer it. Do not make meaning of it.
Instructions. Use the sender and the subject first, then the body. If it fits two buckets, pick the one with a deadline. If it fits none, say so and stop.
Expectations. "My W-2 is wrong" is pay. "Can I carry over PTO" is policy. "Is my spouse covered" is benefits. "I am resigning" is none of the three: flag it for a person.
Feedback. Every Friday, I give it the week's corrected sorts and tell it why I moved them. It tells me what it assumed and what confused it. What we learn goes into this BRIEF. The review is on my calendar.

That is the whole agent. It is deliberately small. It sorts, and it does not do anything else, which is why it can be trusted with the sort.

Now write one

Five boxes, for one task on your desk. Answer them the way you would brief a person, and you will have written the agent.

What you type stays in this browser. There is no server, no account, and nothing is sent anywhere until you decide to send it.

The job title on the req. One line. It names the file you download and heads the BRIEF itself.

The context a new hire needs on day one: the team, the customers, the documents, how things are done here.

One job, one agent. "You sort the shared inbox into three buckets." Not "you handle HR."

The rules, the order of operations, and what needs a person before it goes anywhere.

Three examples of good, three examples of bad. The field almost everyone skips, and the one that matters most.

Show it the final version and tell it why. Ask it what it assumed and what confused it. Update the BRIEF as you both learn. Put the review on the calendar. It can't learn from what it never sees.

Expectations is still empty. It is the field almost everyone skips, and the one that matters most. Three examples of good, three of bad.

When you are ready, confirm and you will see what you just wrote, in plain English and as a document.

Where each part lives

Background
Uploaded knowledge files, plus a short summary at the top of the instructions.
Role
The first sentence of the project instructions, the system prompt, or the skill file.
Instructions
The bulk of the same text block. A rule that only applies to one kind of task goes in a skill that loads for that task.
Expectations
Pasted examples, in the same block. Real ones, not descriptions of one.
Feedback
A review on your calendar, and the BRIEF itself, updated as you both learn. Not a setting. A process.

The tool question

Claude, ChatGPT, Copilot, Gemini. They will stay within five or ten percent of each other and keep passing each other up. Do not pick a vendor. Learn the method. A BRIEF written once runs on all of them.

The Rewire Company is here to help make AI work for you. For more information visit our website or contact Sean Glennon at [email protected].