Plan, Verify, and Explain

For a complex task, improve the prompt by defining the outcome, evidence, constraints, and checks. Ask for a useful explanation of the answer—not a transcript of hidden reasoning.

Use an outcome-first prompt

A strong prompt normally specifies:

1. Goal: the task and the decision or learning outcome it should support.
2. Context: only the information needed to do the work.
3. Evidence: the sources or supplied material the answer may rely on.
4. Success criteria: what a good result must contain.
5. Constraints: audience, scope, tone, length, and prohibited actions.
6. Output: the exact structure you need.
7. Uncertainty: what to do when information is missing or ambiguous.
8. Checks: how the model should test the result before returning it.

Current model guidance likewise recommends stating the outcome and success criteria clearly, then asking the model to validate its work.² Long instructions about every reasoning step can make a prompt more brittle when the exact path does not matter.

Ask for checkable work

The right kind of explanation depends on the task:

Ask for a concise rationale when it helps a reader judge the answer. Do not treat verbosity as rigor.

A reusable prompt pattern

<goal>
Create [deliverable] to support [learning or teaching outcome].
</goal>

<context>
Audience: [learners and level]
Course context: [relevant context]
Available time/resources: [constraints]
</context>

<evidence>
Use only the material below for factual claims:
[paste approved material or provide sources]
</evidence>

<success_criteria>
- [criterion 1]
- [criterion 2]
- [criterion 3]
</success_criteria>

<instructions>
Produce [required format and length].
Distinguish sourced facts from suggestions or inferences.
If essential information is missing, ask up to three questions before drafting.
Do not invent facts, quotations, references, or student information.
</instructions>

<verification>
Before answering, check the draft against every success criterion.
Return the deliverable, followed by a short verification table with one row per
criterion and any uncertainty that remains.
</verification>

Delimiters such as XML-style tags are optional; they mainly help separate instructions from source material in longer prompts. See Separate instructions, sources, and output.

Decompose only where it helps

Break a task into stages when later work depends on earlier results. For example:

1. extract the relevant learning outcomes from an approved course document;
2. propose assessment items mapped to those outcomes;
3. review coverage and difficulty; and
4. revise the set after the educator’s feedback.

This creates useful checkpoints. It does not make an unreliable source reliable, and asking the same model to “double-check” itself is not independent verification.

Make uncertainty actionable

Replace “do not make assumptions” with instructions the model can follow:

List assumptions that materially affect the result. If an essential input is
missing, ask me before continuing. If a minor detail is missing, choose a
reasonable default, label it, and explain how changing it would affect the
answer.

Test and iterate

Evaluate prompts on several representative examples, including difficult and edge cases. Record the model and settings used, the failure patterns, and the human corrections. Re-test important workflows when the model, tool, source material, or task changes.

References & Footnotes