Before you hand a task to an AI agent
Task, context, limits and checks, so the agent does the right thing and breaks nothing.
What it is
A check for the task you give Claude Code, Codex or ChatGPT's agent: a specific result, the files it needs, no sending, paying or deleting without confirmation, a limit and a way to verify the work. It pairs with the prompt "Brief an AI agent on a task."
- AI agents
- Delegation
- Safety
Text to copy
THE TASK
- [ ] The goal is stated in one sentence: why you need this
- [ ] The result is specific: which file, in what format, where to put it
- [ ] There are completion criteria you can check ("a 20-row table, each row with a source link")
CONTEXT
- [ ] The agent has the files and links it needs, not "go find it yourself"
- [ ] It's stated what's already known and what's been tried
- [ ] There's an example of a good result, if you have one
LIMITS
- [ ] You've listed what it must not do without confirmation: send, publish, pay, delete
- [ ] Access is granted only to the folders, mailboxes and services it needs
- [ ] A limit is set: time, number of steps or budget
- [ ] It's told when to stop and ask
CHECKING
- [ ] It's clear how you'll check the result and how long that will take
- [ ] In its report, the agent lists what it did, which sources it used and what it's unsure about
- [ ] The first run is on a small piece of the task, then you scale upLessons that use it
Professions it helps
- Clinical AI implementation specialist
- Software developer
- System administrator
- DevOps engineer
- AI engineer
Showing 5 of 7
More in Checklists
- Before you launch a product · 27 items: payments, analytics, legal basics, support, launch day and a review one week later.
- Data security when working with AI · What not to paste into a chat, which settings to check and how to limit AI agents.
- Onboarding an employee or contractor · What to prepare before the first day, and what to do on day 1, week 1, month 1 and day 90.
Checked: October 2026