The hiring manual for your first AI team members.
This playbook assumes your team already uses ChatGPT or Claude, just without any system around it. If nobody on your team has touched AI yet, start at Chapter 3 instead and come back to Chapter 1 in a month.
Right now, someone on your team is pasting a half-written brief into a chat window. Someone else is on their fourth attempt at the same email. A third person built something genuinely clever last Tuesday, and nobody else will ever see it.
You don't have twelve people using AI. You have twelve people quietly running twelve private agencies.
Each one hired their own freelancer without a brief or standards. And when that person leaves, everything they figured out walks out with them.
This playbook treats AI agents like team members you hire, onboard, brief, review, and sometimes let go. You already know how to run people. Here we'll see how to run agents.
Nothing here requires you to be technical, buy a platform, or hire anyone. Every output in this playbook is a document, an owner, and a review date.
The seven chapters at a glance:
Before you change anything, find out what's already happening.
You can't restructure something you haven't looked at.
Instead of assuming what the team uses AI for, audit it. The audit almost always shows something more you did not know your team was doing with AI.
Treat it like a shadow IT audit with the blame removed. You're collecting data, not policing. Say that out loud before you send anything.
A one-page inventory. Four columns: task, who does it, how often, verdict.
Pick which jobs go to agents and which ones don't.
An agent gets a job, not a topic.
"Content agent" is not a job. "Case study first-drafter" is. The difference shows up in every output you get.
Run every candidate task through four questions. All four have to be yes.
| The question | Why it decides things |
|---|---|
| Does this happen weekly or more often | Below weekly, systematizing costs more than it saves |
| Do you produce several versions of it | Volume is where agents pay off |
| Would two smart people produce roughly the same answer | If judgment varies a lot, keep it human |
| Can someone spot a bad version in under two minutes | If you can't check it, you can't ship it |
This list protects your team more than the agent list does. Say it explicitly:
Three job descriptions plus one stays-human list. One page total.
Everything an agent needs to know before it does anything useful.
Your team is re-explaining the company in every single prompt.
That's the hidden tax on ad-hoc usage. Ten people, ten slightly different versions of who you are, typed fresh every time. A context pack ends that.
Five files. One page each. Boring, specific, and reused everywhere.
| File | What goes in it |
|---|---|
| Who we sell to | Job titles, what they care about, what they've already tried and dropped |
| How we sound | Ten sentences you would write. Ten you would never write |
| What we sell | A plain description with the positioning language stripped out |
| Proof | Real numbers and the customer names you're allowed to use |
| What good looks like | Three pieces you were proud of, three you weren't, one line each on why |
Don't describe your voice. Show it. Ten sentences you'd write, ten you'd never write.
Adjectives like "friendly" and "confident" mean nothing to an agent, and honestly not much to a new copywriter either. Examples work. Patterns get copied.
Five precise files in a shared folder, with a named owner and a review date on the folder.
One brief format. One place where prompts live.
When every person briefs differently, output quality tracks who asked rather than what was needed.
A brief is a brief. It's the same thing you'd hand a freelancer, or an AI.
| Line | What you write |
|---|---|
| Job | Which agent role this goes to |
| Audience | Who reads this |
| Goal | What should happen after they read it |
| Must include | Facts, numbers, links |
| Must avoid | Topics, claims, phrases |
| Done looks like | Length, format, and one example |
Role, then task, then version. For example: "Newsletter drafter — weekly issue — v3." Dull names are findable names.
A one-page brief template and a prompt library with a named librarian.
The chapter that stops quality collapsing when volume goes up.
Reviewing everything kills your time savings. Reviewing nothing is how a wrong number reaches a customer.
The answer is three lanes, and every workflow sits in exactly one of them.
Internal summaries, meeting notes, first-pass internal drafts
Social posts, email variants, ad copy from approved concepts
Anything with a number, a customer name, a claim, a price, or a byline
If it has a number in it, it's red. That single rule prevents most of the damage teams do to themselves.
A one-page lane chart and the amber checklist.
Lanes become approval steps in the workflow. Red-lane tasks can't be closed until the named approver signs off, which makes the gate structural instead of a reminder somebody forgets on a busy day.
Four numbers, checked monthly, on one page.
Output volume alone tells you nothing. More content was never the goal.
| Number | What it tells you |
|---|---|
| Rework rate | Of the AI-assisted work, what share needed a substantial rewrite. Aim for under 30 percent. Climbing means your context pack or your briefs have gone stale |
| Cycle time | Brief to published, compared against your baseline |
| Volume per person | Only meaningful when you read it next to rework rate |
| Reclaimed hours | What people now do that they never had time for before |
Rework rate is the one that matters. If it's climbing, your system is decorative.
Most teams add and never subtract. Write the kill rule down before you need it, because in the moment there's always a reason to keep something alive one more month.
A one-page monthly scorecard.
One page. Deliberately boring. Signed by everyone.
If you haven't written the rule down, the rule is whatever each person guesses it is.
| Section | What it covers |
|---|---|
| What never goes in | Customer data, contracts, unreleased pricing, anything under NDA, employee information |
| Where the work happens | The approved accounts. Free personal accounts are not approved |
| When you say so | Your disclosure rule. Pick one and write it down plainly |
| Who's accountable | The person who ships is accountable. The agent isn't a defense |
| When it goes wrong | Who to tell, in what order, and how fast |
Section one is the section that matters. Read it aloud in a team meeting.
Disclosure expectations vary by market and they're still changing, particularly if you sell into the EU. Check what currently applies to you before you set your rule, and when you're genuinely unsure, disclose.
A signed one-page policy, stored where the work happens.
Print this. Stick it somewhere annoying.
| Chapter | What you produce | Roughly |
|---|---|---|
| 1. Take attendance | A one-page inventory of what's already happening | 2 hours |
| 2. Write the job descriptions | Three agent jobs and one stays-human list | 2 hours |
| 3. Onboard them | Five one-page context files with an owner | 90 minutes |
| 4. Brief them properly | A six-line brief format and a prompt library | 2 hours |
| 5. Review their work | A three-lane chart and the amber checklist | 1 hour |
| 6. Run performance reviews | A one-page monthly scorecard | 1 hour |
| 7. Write the house rules | A signed one-page policy | 45 minutes |
Here's one more thing worth sitting with. The point of restructuring around agents is not to produce more. It's to move your team's hours toward the work only they can do. If your reclaimed hours go into making more of the same thing, you've automated a treadmill.
Decide now what those hours are for. That decision is worth more than any of the seven documents.
Jobs, owners, briefs, review gates, and the numbers all belong next to the work. That's what 5day.io is built for.
Jobs, owners, briefs, review gates, and the numbers — all next to the work. That's what 5day.io is built for.
Start free at 5day.io →