Skip to article

Why Does Every AI Agent Skill Need a Never-Do List?

An AI agent will use whatever you've connected to finish the job. A never-do list draws the lines it can't see. What goes on it and where it lives.

By Jacqueline V. TwillieNovember 6, 20263 min readGuide
Quick answer

Because an AI agent finishes the task with whatever you've connected, and it can't see a line you never drew. A never-do list names those lines: what it must not send, change, or guess. Keep each one next to the job it protects, then test it.

Open the Gmail connector on an AI agent in the platform I build on and the tool list splits in two. Read-only tools sit on one side, six of them, and you can restrict or allow each. Write and delete tools sit on the other: sending emails, creating drafts, forwarding, updating. You're deciding every one of those before the AI agent touches your inbox.

That screen's a never-do list in its most literal form. The skill and the instructions need one too, in plain words.

In my live sessions I put it this way: "I give them guardrails. I tell them what they can do and what they can't do."

Without one, an AI agent finishes the task with whatever's connected. If sending is possible and nobody said no, sending is on the table. If research can go deep and nobody set a limit, a thorough AI agent can dig further than you'd ever be comfortable with.

What does a good never line look like?

Specific and short. You don't need a long list. Four of mine:

  • Never use em dashes, because my readers would feel it
  • Never describe me with titles or framing I did not approve
  • Never send an email, only draft it
  • Never guess a number, name, or date it should ask me about

The first line carries its reason, and the reason does extra work. "No em dashes" covers one punctuation mark. "My readers would feel it" gives the AI agent something to apply to a cousin I never listed, like an over-formal sign-off.

A never line can also carry its own exit. The Chief of Staff file I hand out in my live builds says every reply is "a draft, saved as a Gmail draft, never sent, until I say the words 'send it' in that exact turn." The yes lasts one turn. The next turn, it's a draft again. The file uses the same turn rule for calendar events and posts.

Where does the list live?

Next to the job it protects. In that same file the never lines get a section of their own, six of them: no sending, no creating or moving calendar events on its own, no deleting, no answering invites, no posting or messaging anyone I haven't named, and research connectors that read public information only.

Each AI agent on my team has one job and strict limits on what it can touch. The em dash line lives in my voice skill, with the rest of my style rules.

What do you do before you connect anything?

Write three never lines for your first AI agent: what it must never send, change or guess. Three's plenty to start. An AI agent can ignore its instructions. Ask it to do each of the three and read what comes back. Where the platform has a switch for the tool, flip that too.

More on Real Workflows for Mid-Career Professionals