Ask the AI what it needs
- Industry term
Ask the AI what it needs means describing a task, then asking the assistant which examples, rules, records, or decisions it is missing. You review its list before supplying anything.
A model has read about countless tasks like yours, so it can list the information such tasks usually depend on. It can't see your files or what you already know, so its list is a guess, one that can miss a source your case depends on or ask for things the task never uses.
Where you’ll meet it
You ask your assistant to help write your annual self-review at work, and before it starts, you ask what it needs. It asks for last year's goals, the projects you finished, feedback you've received, and an export of your whole work inbox.
What to do
Give it the goals, the project list, and the feedback, quoted where you have them. Hold back the inbox, which is full of coworkers' private messages. If one thread from it matters to the review, paste that thread alone.
Easy to mix up
| Ask the AI what it needs | Meta-prompting |
|---|---|
| You, adding the material it asked for | The AI, drafting or rewriting it |
Check yourself
Your sign-up page has run slow on Monday mornings for a month. You ask a coding agent what it needs to speed the page up, and it asks for the page's code and its settings. What do you send?
Go deeperA connected agent may fetch what it asked for
In a plain chat, the assistant can use only what you paste or upload, so its list is a request. An agent connected to your email, files, or calendar may fetch the sources it named on its own, within the access its settings allow. With an agent, the review has to happen before it runs.
Asking which missing information would change the answer, and how, gets a list to judge item by item. An item the assistant can't tie to a change in the result is a candidate to leave out. Some assistants ask clarifying questions on their own when a request is unclear; asking for the list makes that step deliberate.