Choose one narrow job

Back to: Build Your First AI Assistant

Define an assistant around one recurring outcome that can be tested.

The idea

“An assistant that helps with everything” cannot have clear instructions or a meaningful test. Start with a narrow job: turn approved meeting notes into a checked action draft, coach a learner through a practice rubric, or help a customer find the correct published resource.

Write a job statement: “For [user], when [trigger], the assistant helps produce [outcome] using [approved inputs], but never [boundary].” Then name what good work looks like and when the assistant must stop.

Confirm the job is frequent enough to matter and bounded enough to review. If it requires unknown private context, unrestricted web actions, or high-stakes professional judgment, reduce the scope or choose another project.

Try it now

1. List three recurring tasks that already have an owner and definition of done.
2. Score each for frequency, input quality, reviewability, error cost, and scope stability.
3. Select one task and write its job statement.
4. Create an in-scope and out-of-scope table with five examples each.
5. Define three success metrics and three hard stop conditions.

ASSISTANT JOB\nFor [USER], when [TRIGGER], help produce [OUTCOME] using [APPROVED INPUTS], but never [BOUNDARY].\n\nIN SCOPE: [5 examples]\nOUT OF SCOPE: [5 examples]\nSUCCESS: [quality, time, correction metrics]\nSTOP AND REFER WHEN: [3 conditions]\nOWNER: [person/role]

Quality check

• The job produces one observable outcome.
• Inputs and users are defined.
• Out-of-scope requests have a safe destination.
• Success includes correction quality, not only speed.