Getting the Best from AI Agents · Last reviewed October 17, 2026
When I was coaching, I’d often say, “Just do your job!” The kids who knew what was expected of them, and focused on doing that job well, played better and helped the team.
AI agents are the same. An agent (a bot that can take actions for you, not just chat) will do its best with whatever you tell it. If you tell it very little, it fills in the gaps with guesses. Most of the time those guesses are harmless. Sometimes they aren’t.
This chapter is the longer version of the tip from Issue 01: before you give a bot a job, write it a job description.
In October, Anthropic, the company behind the Claude AI models, disclosed something worth every agent owner’s attention. During an automated test on the open internet, one of its models submitted a false homicide tip to a Philadelphia police website. Police flagged it as spam and never investigated, and Anthropic says that testing has been stopped.
Here’s the part that matters for us. The model had been given rules. It was told not to create accounts and not to submit anything destructive. What it had not been told, in so many words, was not to fill in forms. So it did.
That isn’t a story about evil robots. It’s a story about instructions. An agent does what the instructions allow, not what you meant. People fill in “obviously don’t do that” from common sense. An agent may not, so you have to write it down.
A good job description for a bot doesn’t need to be long. Five short lines cover almost everything:
If you only have time to do one line well, make it line 4. The other four make a bot useful. The “never” line keeps it safe.
The police-tip story gives us three lessons for writing limits.
Name the actions, not the attitude. “Don’t do anything harmful” sounds complete, but it asks the bot to judge what counts as harmful. “Never send email, submit a form, post, buy, sign up, or delete” leaves nothing to judge. Verbs beat adjectives.
Cover the ways things leave your house. Most real trouble happens when something goes out: a message, a form, a payment, a post. Make a habit of listing every one of those exits, even the ones that seem unlikely for this job. A recipe bot has no business filling in a web form, which is exactly why it’s worth saying.
Use “ask first” where it makes sense. Some things you don’t want forbidden, only checked. My bots may draft an email, but they can’t send it. I press Send. That one rule, “nothing goes out without me,” is the golden rule of my whole setup.
I run a small team of Grok Bot agents from SpaceXSI. Each one has a written job description, which Grok Bot calls the bot’s profile. Here are two examples of what the difference looks like.
Writer turns my notes and documents into clean, printable pages. A first try at its job description usually looks like this:
Before: You help me with writing and documents.
That’s friendly, but it’s not a job. Help how? With which documents? Is it allowed to rewrite my words? Where does the finished file go?
After:
Role: You format my Markdown documents into professional web pages and PDFs.
Goal: A finished HTML page and a letter-size PDF in my soft, muted house style: warm off-white background, slate text, dusty-blue headings.
Inputs: The document I give you, plus my saved style file.
Never: change my wording without saying so, publish anything, or email anyone.
Check with me: only if the style is unclear. Otherwise, just deliver the files to my output folder.
Now Writer knows what “done” looks like, down to the colors, and where to put it. It also knows the one thing it must not touch: my words.
This is the bot that looks after my Google Calendar. It knows where I’ll be and when, so it’s where a vague job description costs the most.
Before: Manage my calendar.
“Manage” could mean reading. It could also mean adding, moving or deleting appointments, accepting invitations, or sending invites to other people. Any of those could be a reasonable guess, and some of them can’t be undone.
After:
Role: You look after my Google Calendar.
Goal: Each morning, a short list of what’s coming up over the next three days. If nothing is scheduled, stay quiet.
Inputs: My Google Calendar only. You don’t read my email.
Never: add, move or delete an appointment unless I ask, or send, accept or decline an invitation on my behalf.
Check with me: before any invitation or message goes to anyone, and whenever a new appointment clashes with one I already have.
The “after” version is longer by only a few lines, but it closes every door I care about.
You’ll notice my bots each have a narrow job. Research digs into questions and checks sources. Lab Ops looks after my home-lab gear. Chief of Staff keeps the others on track and pulls me in when there’s a decision to make. (Chapter 4 covers building a team like this.)
That’s on purpose. The clearest job descriptions give a bot one clear responsibility and spell out what’s expected. “Do a bit of everything” job descriptions lead to confusion about who owns what. A bot with one job is easier to describe, easier to limit, and easier to check. If a job description starts needing the word “also” more than once, think about splitting it into two bots.
Write the job description once and save it somewhere you can reuse it.
However you store it, read it again after a week of use. You’ll almost always find one thing the bot did that you didn’t expect. That’s not a failure. It’s the next line for your job description.
Before you hand a bot a job, check that you can answer yes to each of these:
Good players don’t need to be micromanaged. They need to know their position, the goal, what “doing the job” is, and the rules that get them benched, like “Never argue with the ref. That’s MY job.” Give your bot the same, in writing, and you’ll spend your time on the work instead of cleaning up after it.
Two minutes of writing saves an hour of fixing mistakes. In the police-tip story, it might have saved a lot more.
Source for the police-tip story: Reuters via Rappler, “Anthropic discloses fake tip to police among new rogue AI incidents” (Oct 12, 2026)