Lesson 02: Write the Ask

Lesson objectives:

• Build a clear ask from four parts: task, requirements, boundaries, and acceptance criteria.
• Spot which part is missing in a vague request.
• Rewrite one vague wish into a four-part ask.

Prerequisites: Lesson 01 | Previous << 01 Worker not talker | Next 03 Check the result >>

The result is only as good as the instructions you forgot to give

Agents do not read minds. They read your prompt. When the prompt is vague, they fill the gap with the most common pattern they have seen in training. That pattern is rarely your exact situation. The fix is not to write more words. It is to give the agent the four pieces of information it actually needs¹².

Explanation

A clear ask has four parts. You can remember them with the phrase "What do I want, how should it look, what should it avoid, and how will I judge it?"

1. Task — the one action the agent should perform. One sentence. No hidden steps.
2. Requirements — the properties the output must have. Format, length, tone, sources, structure.
3. Boundaries — what the agent should not do. This stops scope creep and unwanted side trips.
4. Acceptance criteria — the checklist you will use to decide if the result is good enough (the same checklist you will verify against later).

For example, a weekend trip ask could be: "Plan a weekend trip within a $300 budget. Return a table with two options, each showing total cost, travel time, and one activity. Do not include flights. Acceptance criteria: both options are under $300, each has a travel time, and each names one activity."

These four parts are not a gimmick. They are the same pieces that OpenAI and Anthropic guides recommend: be specific about the desired output, provide context, and define what success looks like¹².

A vague wish like "make my résumé better" gives none of these. The agent has to guess what "better" means. A clear ask says: "Rewrite my résumé bullet points for a product-manager role. Keep each bullet to one line, start each with a verb, and remove jargon. Acceptance criteria: every bullet is one line, starts with a verb, and mentions a measurable outcome."

The diagram shows that the four parts are sequential. You decide the action first, then shape the output, then fence off unwanted behavior, and finally define the check.

Worked example (follow along)

Marcus wants to compare two laptops for his daughter. He starts with a vague wish:

"Which laptop should I buy?"

The agent will either ask a lot of follow-up questions or give a generic list. Marcus does not want either. He rewrites it with the four parts:

Task: Read the two product pages I will paste below and compare the two laptops. Requirements: Return a table with three rows: CPU, RAM, and price. Each cell should have one value or "not listed." Keep the tone neutral. Boundaries: Do not recommend which one to buy. Do not mention laptops outside these two. Acceptance criteria: Every cell has a value or "not listed," and the table has exactly the three rows I asked for.

Now Marcus gets a table he can verify against the product pages. He also avoided the agent's urge to make the decision for him.

Your turn (faded example)

Fill in the missing parts of this ask.

Task: __________ the customer feedback spreadsheet I will attach. Requirements: Group feedback into three categories: praise, bug, and feature request. Count the items in each category. Boundaries: Do not __________ any customer names. Acceptance criteria: The output has three rows, one per category, and a number in each row that I can check against the spreadsheet.

Answer: Task: "Read and categorize"; Boundaries: "include" or "reveal."

Summary + what's next

A clear ask is not longer. It is structured. It tells the agent what to do, what the output should look like, what to avoid, and how you will judge it. In the next lesson, you will use that acceptance criteria as a checklist to verify the agent's result.