AI Learning Hub › Learn AI › Self-Paced Courses › Introduction to AI › Intermediate › Prompting Is Task Design

Level 3 of 4 · Intermediate · Lesson 1 of 6

Prompting Is Task Design

Designing AI tasks with clear goals, instructions, constraints, context, examples, and output formats.

About 15 minutes

Skip to lesson

By the end of this lesson you should be able to

  • Turn a vague request into a bounded task with a goal, an audience, supplied evidence, constraints, an output form, and acceptance criteria
  • Decide when a missing detail should be supplied, left as a visible placeholder, or asked about, rather than filled in by the assistant
  • Revise a request by comparing the result with your criteria, not by rephrasing until an answer looks right

Why this matters

The Beginner level explained why an assistant’s output can be fluent and still wrong. This level is about what you can do about it, and the first thing you control is the task itself. An assistant cannot see what you know but did not write down. Asked to “write an email about the workshop,” it may guess at the purpose, the reader, the facts, and the length; it may ask questions or leave placeholders; or it may return a draft you cannot use. A clear brief reduces that ambiguity. It does not guarantee the assistant will follow it, which is why the result still gets checked.

So the question for this lesson is: what must be clear before I ask an assistant to do this work? Think of a prompt as a brief for a capable collaborator who has none of your background.

Six things to settle before you ask

Goal
What the output is for: the decision, action, or understanding it should support. “So that interested students register,” not just “an email.”
Audience
Who will read it and what they already know. A note to first-year students and a note to faculty colleagues differ in more than tone.
Supplied evidence
The material the answer must rest on, such as a pasted brief or your own notes, and whether anything outside it may be used. Naming a reading does not give the assistant its contents: supply the text you are permitted to use, or confirm that the tool can actually retrieve it.
Constraints
Limits the result must respect: length, what to leave out, what must not be invented, and any course or office rules that apply.
Output form
The shape of the result: a subject line and body, three bullet points, a table with named columns.
Acceptance criteria
How you will judge the result, decided before you see it. “Every date matches the brief.” “Under 150 words.” “Anything unconfirmed appears as a placeholder.”

Acceptance criteria are easy to skip, and without them you cannot tell whether a revision is better. Anthropic’s prompting documentation starts from the same place: it assumes you already have “a clear definition of the success criteria” and some way to test against them before you start refining a prompt. (Anthropic, “Prompt engineering overview” (opens in a new tab); provider documentation.)

Ask for what you can check

Part of a good output form is asking for material you can verify: the source and location for each factual claim, any assumptions stated explicitly, calculations shown with their inputs, and a short explanation of the method. That is different from asking a system to reveal its hidden reasoning. As Beginner Lesson 4 noted, a displayed explanation is something to assess, not a guaranteed record of how the answer was produced.

Where C.R.A.F.T. and roles fit

If you learned the library’s earlier memory aid, C.R.A.F.T. (Context, Role, Action, Format, Tone), it still works as a drafting checklist; the disclosure below maps it onto the six elements. One caution applies to its Role: an instruction such as “act as an experienced appellate clerk” can set a perspective, vocabulary, and register. Adopting a role does not give the system a clerk’s credentials, authority, or access to sources, and it does not establish that the output is reliable. If a role instruction seems to help, check that on the task itself.

Optional: how C.R.A.F.T. maps onto the six elements

Context covers your background and supplied evidence. Action is the operation you want (draft, summarize, compare), which is not the same as the goal, what the work should accomplish; state both. Format is the output form. Tone belongs with the audience. Role can set perspective, as above. C.R.A.F.T. has no letter for constraints or acceptance criteria. You could write them into the context, but naming them separately makes them harder to forget and easier to check. Treat C.R.A.F.T. as a checklist for writing a brief, not as a guarantee about the result.

Optional: five common gaps, and what fills them

These five come from examples of weak legal prompts in the Hub’s earlier introductory AI course. Each is a missing element, not a missing magic word.

  • The vague request. “Research employment law for me” has no jurisdiction, no issue, and no output form. Name the question, the jurisdiction, and what you need back.
  • Assumed knowledge. “What was the holding in the Johnson case?” names no citation, court, or date, and many cases share a name. Supply the case, or ask for help finding it, not for its holding.
  • The kitchen sink. Asking for research, a brief, and a client memo at once leaves each task underspecified and gives you no single result to check. Split it into separate tasks.
  • The unstated standard. “Review this contract and tell me if it’s good” leaves out whose interests, which clauses, and which law. State the perspective and the criteria.
  • The fact-free draft. “Write a demand letter” with no facts, claim, or desired outcome invites a template the system fills with guesses. Supply the facts and the aim, or do not ask yet.

Supply it, mark it, or ask

When a detail is missing, you have three honest options:

  • Supply it, if you know it or can find it.
  • Mark it: tell the assistant to leave a visible placeholder, such as [ROOM TO CONFIRM], exactly where the detail belongs.
  • Ask: have the assistant list what it needs before it drafts. “Before drafting, list any information you need that is not in the brief. Do not guess it.”

A useful rule: if a wrong guess would change a fact someone relies on (a date, a room, a deadline, a legal standard), do not let the assistant fill it. If the gap is a matter of style, let the draft propose something and decide for yourself.

Worked example: from a vague request to a bounded task

A constructed example. The workshop, people, and assistant output are fictional.

A staff member has this workshop brief and needs an email to students.

Workshop brief. Title: Finding Legislative History. Presenter: Priya Natarajan, reference librarian. Date: Wednesday, February 10, 12:15–1:15 p.m. Room: to be confirmed. For: 2L and 3L students writing seminar papers. Capacity: 20. Registration required; link to follow. Bring a laptop. No food provided.

The first request was: “Write an email about the legislative history workshop.” Here is the same request as a bounded task.

A bounded task

Goal: Draft an email inviting students to the workshop described in the brief below, so that interested students register.

Audience: 2L and 3L students writing seminar papers. They know what legislative history is but may not know how to research it.

Evidence: Use only the workshop brief below. Do not add facts that are not in it.

Constraints: Under 150 words. The room and the registration link are not confirmed: write [ROOM TO CONFIRM] and [REGISTRATION LINK] where they belong. Do not mention food, recording, or credit.

Output: A subject line, then the email body in plain text.

Before you finish, check: every date, time, name, and number matches the brief; both placeholders appear; nothing is promised that the brief does not say.

[Workshop brief pasted here]

What each addition does:

  • The goal tells the assistant the email should lead to registration, so the draft should make the registration step prominent.
  • The evidence line instructs the assistant to use only the brief, so every fact in the draft should be checkable against it.
  • The constraints say how to handle the two unconfirmed details and which extras to leave out. Like the evidence line, they are instructions, not barriers; the draft still has to be checked against them.
  • The criteria are the checklist you will use yourself when the draft comes back.

Suppose the draft comes back like this (constructed):

Subject: Research legislative history for your seminar paper — Feb. 10. … Join reference librarian Priya Natarajan on Wednesday, February 10, 12:15–1:15 p.m., in the Library Classroom. Space is limited to 20 students; register at [REGISTRATION LINK]. Bring a laptop. …

Checked against the criteria, one failure is visible: the room is not confirmed, and “the Library Classroom” appears nowhere in the brief, even though the task said to use a placeholder. (Only part of the draft is shown; the omitted text would need the same check.) The useful response is specific: “The room is unconfirmed. Replace ‘the Library Classroom’ with [ROOM TO CONFIRM].” Rewording the whole request, or asking again and again until a version happens to look right, would not tell you what went wrong.

Iterate against the criteria

  1. Compare the result with each criterion.
  2. Name the specific miss, and whether it came from the task, the evidence, or the output.
  3. Fix that: add the missing information, tighten the constraint, or correct the draft yourself.
  4. Stop when the criteria are met, or when the problem is not a prompting problem.

Not every problem yields to a better prompt. If the information does not exist yet, if the tool cannot reach the source, or if the task needs judgment you have to exercise yourself, more rewording will not help. And rephrasing until the assistant agrees with you is looking for agreement, not quality (Beginner Lesson 3).

Try it: bound an underspecified request

A constructed scenario. The course, its policy, and the materials are fictional; the policy below is not a Stanford policy.

A 1L wants help preparing for a midterm and has typed: “Make me a study guide for the midterm.” Here is what the student has:

  • From the syllabus: “The midterm covers Units 1–4: offer and acceptance, consideration, promissory estoppel, and the statute of frauds. It is closed-book: three essay questions in 90 minutes.”
  • The course AI policy (fictional): “You may use AI tools for self-study. You may not use them during the exam.”
  • The student’s own outline of Units 1–4, which is complete for Units 1–3 and has only headings for Unit 4.
  1. Rewrite the request as a bounded task, using the six elements.
  2. Which gap should the assistant not fill on its own, and how will your task handle it?
  3. Write two acceptance criteria you would check the result against.
A model revision and discussion

This is one well-bounded version. Others could work as well, as long as they settle the same questions.

Model revision

Goal: Help me prepare for a closed-book Contracts midterm: three essay questions in 90 minutes on Units 1-4 (offer and acceptance, consideration, promissory estoppel, statute of frauds).

Audience: me, a 1L who has taken notes on all four units.

Evidence: Work only from my outline, pasted below. My Unit 4 notes are headings only. Do not fill them in from general knowledge; mark that section [UNIT 4 NOTES INCOMPLETE] and tell me what to review.

Output: For each of Units 1-3, the rules as my outline states them, then one short essay-style practice question. Put no model answers in this response; I will ask for feedback on my own answers separately.

Check before you finish: every rule statement can be found in my outline; Unit 4 is flagged rather than written; there are exactly three practice questions.

[My outline pasted here]
  1. Why these choices. The goal names the exam format, so the practice matches it. The evidence line keeps the guide tied to the student’s own notes, which the student can check. That makes the guide checkable against the outline, not the outline correct: rule statements still need checking against the assigned readings. Holding back model answers keeps the exercise about the student’s learning.
  2. The gap is Unit 4. A guide that quietly filled it in would look complete and could not be checked against anything the student has. Marking it, and returning to the readings, is the honest option. The student could also add the Unit 4 readings to the evidence.
  3. Criteria could include: every rule statement appears in the outline; Unit 4 is flagged, not written; three practice questions, one per unit covered. Anything similar that you can check without trusting the assistant works.

The fictional policy permits self-study, so the use is allowed as described. A real course’s policy, or the SLS default where a course has none, is covered in Beginner Lesson 5.

Review starts with your criteria. Write down how you will judge the result before you ask. When the output comes back, check it against those criteria, not against how convincing it sounds, and fix the specific thing that failed.

Sources and further practice