AI Agent Playbook
One agent, one job, and nothing goes out until you say yes.
One agent, one job, and nothing goes out until you say yes.
This playbook is how you build your first AI agent and hand it a single piece of your business, the way one agent can own proposals, another can own invoicing, and another can answer client questions once a job is done. It runs on Claude and Cowork from a plain conversation, no connectors needed, and the setup prompt at the bottom works in any AI.
The problem
You already do the same job every week, the kind with a clear start and a clear finish: quote a proposal, send an invoice at the right milestone, answer the same question from a client or a tech in the field. Modeled at 30 minutes a task, five times a week, that is over two hours every week spent on work that follows the same pattern each time.
The real cost is not the time. It is what happens when that job depends on you being in the room. A proposal waits until you sit down on a Sunday night. An invoice slips a week because the milestone came and went while you were on a job site. A client texts at 9pm and gets an answer the next morning, or no answer at all.
The fix is not hiring for one task, and it is not a tool that tries to do everything. It is giving that one task an owner who never forgets it, works only from what your business actually knows, and never sends anything without you.
What you get
How it works
Pick one job. Name the single task that eats the most repeated hours, the one with a clear start and a clear finish: a proposal, an invoice, a recurring client question. Not "help with sales." One job.
Write the lane. One sentence, in plain language: what this agent owns and what it produces. Read it back and check it only describes one job. If it takes two sentences to explain, it is two agents.
Write the limits. Name what it never does without you. At minimum: it never sends, never spends, and never promises anything to a real person. If the job touches money or a client relationship, the limits matter more than the lane.
Connect it to the Business Brain. Point it at the one place that holds the facts it needs: your price list, your contract terms, your past job documentation, your process notes. It reads from there and only from there. An agent that guesses a price or invents a policy is worse than no agent.
Build the approve-before-send gate. Every output it produces sits as a draft. You read it, you fix what is wrong, you send it yourself, or you tell the agent to send it once you trust the pattern. This gate never turns off for money, for a promise, or for anything that leaves your business and reaches a real person.
Test it on a job you already finished. Feed it the real inputs from a job you closed last month and compare its draft to what you actually sent. This is where you find out if the lane is too wide, the limits are too loose, or the Business Brain connection is missing a fact it needed.
Run it live, then watch the first ten. Put it on the next real job and read every draft closely for the first ten times. Most of what goes wrong shows up in that window: a fact it should have asked for instead of guessing, a tone that does not sound like you, a case the lane did not anticipate.
Add the next agent only once the first is clean. One agent running well beats three agents running rough. When the first one has gone ten drafts without a correction, pick the next job and start again from step one.
Know the ceiling. One agent on one job, reading from a folder, is real leverage and it covers most owners for months. The day you want three or four agents coordinating with each other, one of them reporting to another, or the whole set running against your live inbox and calendar instead of files you paste in, you are past what one owner wiring this up alone can hold together cleanly. That is agent-roster territory, and it is the work we do at Obsessed AI for trade and service businesses. When you get there, obsessedai.co.
Example output
A Job Card written for a five-person electrical company, filled out in about ten minutes:
Agent: Invoice Agent Job: Prepares the client invoice at each of the four project milestones Reads from: Business Brain > Contracts, Business Brain > Milestone Schedule Never: sends an invoice, changes an amount, waives a fee or a late charge Approve-before-send gate: every invoice sits as a draft in the shared folder until the owner opens it and clicks send Test run: re-ran the last two milestone invoices from the March panel upgrade job, both drafts matched what was actually billed within a rounding difference on labor hours
From the next job forward, the milestone invoice is a five-minute review instead of a from-scratch write.
How to trigger it
Works with
The core prompt
Paste this into a new project in the AI you use. It builds your first agent's Job Card, then stays in the project's custom instructions as that agent's operating manual.
You are helping me build my first AI agent: one that owns exactly one
job in my business and never acts on its own past the edge of that job.
Work through these steps in order. Ask questions in batches, never one
at a time, and show me everything before treating it as final.
STEP 0 - PICK THE JOB
Ask me: what is the single task I do most often that follows the same
pattern every time, with a clear start and a clear finish? Push back if
my answer covers more than one job ("help with sales" is not one job,
"write the proposal once I've walked a site" is).
STEP 1 - THE LANE
Help me write one sentence describing exactly what this agent does and
produces. Read it back to me. If it takes more than one sentence, stop
and tell me it is actually two agents.
STEP 2 - THE LIMITS
Ask me: what should this agent never do without me? At minimum confirm
it never sends anything to a real person, never spends or commits money,
and never promises something on my behalf. Write these down as a short
list.
STEP 3 - THE BUSINESS BRAIN CONNECTION
Ask me where the facts this job needs already live: a price list, a
contract, past job documentation, a process doc. If I do not have one
yet, tell me plainly that this agent should not guess a price, a policy,
or a fact it was not given, and that placeholder mode (clearly marked
placeholders instead of invented numbers) is the honest fallback until I
build that file.
STEP 4 - THE APPROVE-BEFORE-SEND GATE
State this back to me as a rule the agent will follow forever: every
output is a draft. Nothing it produces reaches a client, a vendor, or my
bank without me reading it and saying yes first, no exceptions for
speed or convenience.
STEP 5 - THE TEST RUN
Ask me for one real job I already finished. Have me give you the same
inputs that job actually had, produce the draft the agent would have
produced, and then ask me how it compares to what I actually sent. Fix
the lane, the limits, or the Business Brain connection based on what
that comparison shows.
STEP 6 - WRITE THE JOB CARD
Produce a short, saved summary: Agent name, Job, Reads from, Never,
Approve-before-send gate, Test run result. This is the file I keep and
update as the agent's job, limits, and Business Brain source change.
End by telling me: "Run this agent on your next real job and read every
draft closely for the first ten times. Only after it runs ten drafts
clean should you come back and build a second one."This agent is only as sharp as what it reads from. Business Brain is the one-folder setup that gives it real facts to work from instead of a blank chat, and it is worth building first if you have not already.
This playbook builds the structure. Your job, your limits, and your Business Brain make it yours. It teaches you to wire up one agent by hand, on one job, with a human approval gate that never turns off. A ready-made agent roster, with several agents coordinating and a chief of staff routing between them, is past what this playbook alone covers, and it is a different level of build.