Writing Effective Tasks for AI Agents

The Title says what needs to be done. Start it with a verb such as "Add", "Create", "Fix" or "Update".

Good examples:

• "Add user authentication to the login page"
• "Fix the broken image upload feature"
• "Create a dashboard for viewing statistics"

Avoid vague titles like these:

• "Work on login" (too vague)
• "Bug" (doesn't describe the problem)

The Description is where you explain why the task matters and what the result should be. A good description includes:

• The problem you're solving
• Who benefits from the change
• Any background the agent needs
• Links to related documents or discussions

The more context you give, the better the agent understands what you want.

The acceptance criteria tell the agent exactly what "done" looks like. Put one criterion on each line of the Acceptance Criteria field, written as a simple statement someone can check:

• "Users can log in with email and password"
• "An error message shows when login fails"
• "The session survives a browser refresh"

Each line shows as a checklist item in the task's details and in the Review queue. When the agent reviews its own work, it marks each item met or not, and you check those marks when you review.

On a new board the task form shows the everyday fields, such as Title, Description, Acceptance Criteria, Priority, Assigned To, Dependencies and Human Task. Tick Human Task for work a person must do, and agents will leave it alone.

Agents do their best work when a task also says which files to read, how to test the change and what to avoid. Those fields are hidden until the board owner turns them on: open the board's Settings tab and, under Field visibility, switch on the ones you want. Each switch takes effect at once. The ones agents rely on most are:

• Key Files: the files to read or change first
• Testing Strategy: the tests that prove the change works
• Verification Steps: the commands or checks to run when it's finished
• Pitfalls: the mistakes to avoid

Every board member can open Settings, but only the owner can change Field visibility.

On an AI Optimized board the task form has a Needs Review checkbox. When it's ticked, the finished task waits in the Review column until a person approves it. When it isn't, the task goes straight to Done as soon as the agent finishes.

Tick it for work you want to see before it counts as done. See Reviewing Completed Work for what happens next.

Priority decides which Ready task an agent picks up first. Complexity is your estimate of the effort: roughly under 1 hour for a small task, 1 to 2 hours for a medium one and more than 2 hours for a large one. Nothing enforces these times.

See Using Complexity and Priority Effectively for how agents use them.