Quick test. If you hired an assistant tomorrow, could you write their job down in one sentence? What they do, what they're allowed to use, and what they have to check with you first?
Most of us could do that for a person. Almost nobody does it for an AI agent. We pick a tool, connect it to everything, and tell it to "help with marketing." Then it does something strange and we decide agents aren't ready yet.
A few weeks ago James Pasmantier, founder of NimbleDraft, ran a mentor session for our community, and my biggest takeaway was this: an agent with a vague job has to guess, and you can't check a guess. Give it one clear job and you can tell in ten seconds whether it did it right.
James P. showed us the one-page plan he fills in before he builds anything. He calls it the Agent Blueprint. It's eight fields in plain English, and it works the same whether you end up on Hermes, Claude, ChatGPT, or whatever launches next month. He ran it on the smallest job he could think of: a message on his phone each morning with up to three things that deserve his attention today, and why.
Here it is, field by field, with his answers. James P. has since turned the session into a five-part course where you build a working agent with him, step by step. It's live now for community members, and the details are at the bottom.
First, check the job deserves an agent
Before the Blueprint, James P. asks whether the job needs an agent at all.
A one-off question? Ask ChatGPT or Claude and move on.
The same fixed steps every time? That's a regular automation.
A job that repeats, runs on information you can hand over, and produces something you can check? That one's worth it.
His example passes the test. It's Tuesday morning and he has three open tasks:
A customer quote, due today at noon, ahead of a 3 PM client call
A kickoff plan, due tomorrow at 10 AM, with scope and owners still unsettled
A supplier contract that expires Friday, with no date set for the renewal decision
He wants an assistant to look at those and tell him where to start. Here's how he wrote that job down.
1. Give it a role
Who does this assistant help, and in what role? James P. calls his a chief of staff, and named it Eve. The role tells the agent how to look at the work.
Rule: A title is a point of view. It doesn't give the agent any ability or access on its own.
2. Give it one job
His is one sentence: review my open tasks and explain what needs my attention today, and why.
Rule: If you can't say the job in one sentence, it's two jobs. Pick one.
3. Describe the finished result
He wants up to three recommended actions. Each one has a task name he recognizes, the deadline, why it matters, and a suggested next step. The "up to" is deliberate. If only one thing matters today, he wants one. "Please don't overwhelm me," as he put it, and don't invent extra work to fill the list.
Rule: Write down what a good result looks like before the agent produces anything. That's your answer key.
4. Decide what starts it
There are three options: you ask for it, it runs on a schedule, or it checks every few minutes for something (a certain kind of email landing, say). While he was testing, James P. asked for the brief himself each time. It only went on a weekday-morning schedule once he was happy with what came back.
Rule: Run it by hand until you trust it. Schedule it after.
5. Name what it's allowed to use
His agent gets two things: his open task list, and a short file of priority rules. The list says what's outstanding. The rules say how to rank it: confirmed deadlines first, then decisions that are still open, then anything with missing information.
Rule: The agent works only from what you hand it. Two inputs is plenty for a first job.
6. List what it can do
Read those two inputs. Write the brief into the chat. That's the whole list. The tasks themselves stay untouched.
Rule: Start with "read" and "write to me." Add nothing else yet.
7. Set the limits
James P. still sends the customer quote himself, settles the kickoff scope himself, and makes the supplier decision himself. The agent can't make commitments for him or spend money. When it drafts outreach, his instruction is "don't send it unless I've read it first."
He added one thing I hadn't thought about. Typing "don't change my tasks" into the instructions isn't enough. Where the tool lets you, give the agent read-only access, then test that it really can't edit anything.
Rule: Back every written limit with an actual permission setting, and test it.
8. Pick where the result lands
His brief arrives in Telegram each morning. Slack works too. It's short: task names he recognizes, deadlines, the next step, and anything that needs clearing up.
Rule: Send it to the place you already check first thing.
What came back
Here's the brief those eight answers produced:
Customer quote. Due today at noon. Review and send it before the 3 PM call.
Kickoff plan. Due tomorrow at 10 AM. Settle scope and owners so the team can prepare.
Supplier contract. Expires Friday, but nobody has set a deadline for the decision. Confirm that date before reviewing options.
Look at the third one. The agent didn't guess a deadline. It told him one was missing. That only happened because field 3 said to flag gaps and field 7 said not to act alone.
Three more things before you switch it on
What should it remember? Your priority rules and how you like the brief laid out. Not yesterday's task list. That gets read fresh every run.
What happens when it breaks? If it can't reach the task list, it should say so and ask for help.
How will you test it? Use a small example where you already know the right answer. James P.: "I've seen a lot of people try to rush into automation before they've actually done it themselves manually."
Steal this sentence
If eight fields feels like a lot for today, James P. has a one-sentence version. Fill in the brackets for a task you keep coming back to:
When [trigger] happens, I want my assistant to use [approved information] to prepare [result], so I can [next action or decision]. It must ask me before [action needing approval].
That's your first job description. It'll work in whatever tool you pick.
Where this goes
The morning brief is job one. The same setup grows from there, one written job at a time:
Start with a job that only reads. A daily brief changes nothing, so a mistake costs you nothing.
Then let it prepare the work. Once you trust the brief, have it do the legwork behind a task. For outreach, that's a ranked list of people to contact and a first draft of each message.
Feed it context as you go. The more it knows about your business and who you already know, the better its suggestions get.
Keep the send button. It drafts. You read, adjust and send.
One more thing on that second step. If the people you reach out to are local businesses, finding the right ones is the slow part. That's the half I built boringfunnels for: type a trade and a city, and you get a ranked list of local businesses, what each one is losing customers on, and an email your agent can start from. Wanna see what it builds? Here's a whole funnel for a local roofer (a sample business).
Build it with James P.
If you'd rather build one than read about it, James P. turned the session into a five-part course: AI Answer to AI Operator. It's five recorded sessions, just under two hours in total, and it's live now in the Boring Marketing community. Members can unlock it for $199.
By the end, you have an agent that drafts your follow-up messages every weekday morning. It works from a short brief and a list of people you choose, posts up to three drafts in Slack at 8 AM, and never sends anything. You read them, fix what needs fixing, and send them yourself.
Here's what each session covers:
Start Hermes and connect it to Slack. Your agent replies to you in a private channel.
Give it context and a skill. It learns a short business brief and an outreach skill, then writes its first practice drafts.
Put it to work from your notes. It prepares a pack of up to three drafts, and you check each one against the notes.
Let it run on its own. The pack arrives every weekday at 8 AM. You also practice pausing it, and fixing it when a run fails.
Make it yours. Change a rule to fit your business, add your own people, and check the whole system end to end.
Every session comes with a beginner guide and copy-ready prompts, so you're never working it out from scratch. You'll need a ChatGPT account, Claude Code, and a Slack workspace you can add an app to.
Not a member yet? Join The Boring Marketing Community first. Members also get the free session recording and the Agent Blueprint PDF.
Miscellaneous vibes
How I'd build a one-person million dollar agency with AI agents: one niche, three productized tiers, and a delivery team of agents that asks before it acts
My AI coding tools had no idea how to build a good funnel, so I had my agents study the funnels instead
Skip the pitch deck. Send a local business its new website, already built, and let the site do the selling
Before your Black Friday ads go live, buy from your own store on your phone and check these 4 screens
Greg Isenberg: watch AI agents ship 100+ static ads and make UGC videos for $4 that used to cost $500
Idea Browser's reminder that nobody cares about your product until it entertains them, makes them money, raises their status or gets them promoted
Google's AI Mode now links straight to booking home services through Angi, Thumbtack, Housecall Pro and Jobber, with ServiceTitan coming soon. Google hasn't said how it picks who gets recommended
Google is rolling out AI-powered messaging on Business Profiles to more businesses. An AI agent answers routine customer texts and hands off what it can't, so update your hours and services before you switch it on
Keep it boring
Pick one task you keep coming back to and fill in that sentence today. Then hit reply and send it to me. I read every response.
Wanna see a funnel that's selling right now, ad to last email? Get this week's free funnel.
Wanna build alongside 2,500+ marketers? Join The Boring Marketing Community.
Know someone who wants AI doing real work for them? Forward this to them.
Keep it boring,
James