Ask a capable new hire to help with "the project" and they'll find something to do. Maybe the right thing. Maybe they'll redo the schedule you were happy with.
Bots are similar, with one difference: they're very eager and very fast. A clear boundary helps.
Scope in ISO
ISO asks an organization to define the scope of its quality management system: what it covers, and what it doesn't. If something is left out, you say so and why.
It's the first thing an auditor looks at. You can't check a system if nobody knows where its edges are.
The bot version: an anti-job list
For a bot, scope is its job, plus a short list of what it doesn't do. Near the top of its instructions:
I draft replies to customer emails. > I don't send them. > I don't promise delivery dates. > I don't discuss pricing. I flag those for the owner.
That's it. Four lines, and a lot of trouble avoided.
Why "isn't" matters as much as "is"
| Without an anti-job list | With one |
|---|---|
| Bot "helpfully" does nearby tasks | Bot stays on its job |
| Two bots both do the same thing | Each knows its edges |
| You find out about the overreach later | The bot flags it and asks |
| Hard to review: what was it supposed to do? | Easy to review |
A model is trained to be helpful. Left alone, it'll often help with more than you asked. The anti-job list gives it permission to say "that's not mine."
What goes on the list
- Actions it never takes: send, spend, post, delete.
- Topics it doesn't handle: legal, pricing, hiring.
- Other bots' jobs: "Publishing is the publishing bot's job."
- Decisions it doesn't make: "I don't change priorities."
Keep it short. Five to ten lines. A long list gets ignored, by people and bots alike.
A real example
In our review of 90 published marketplace templates, most had a short, explicit anti-job list near the top ("I don't…") (good-practice.md). Our own AN Creation skills had these limits spread through the text, which made them easier to miss. The fix was to gather them in one place.
Where to send what's out of scope
A good anti-job list also says what to do instead. "I don't discuss pricing" is good. "I don't discuss pricing; I flag it for the owner" is better. The bot knows its edge and the next step.
What this means for you
- Write one line for what the bot does.
- Write five to ten lines for what it doesn't.
- Put them near the top.
- Say where out-of-scope work goes.
A clear edge isn't a limit on a good bot. It's what lets you trust it inside the line.