Act as a prompt architect. Help me build one dependable prompt for [TASK TO COMPLETE] using [AI TOOL OR MODEL]. The prompt will be used by [INTENDED USER] and should produce [DESIRED OUTPUT]. First, identify the real job to be done, the decisions the model must make, and the information it cannot safely assume. Then create a finished prompt that I can copy directly into the AI tool. Use this structure in the finished prompt: 1. Role and task: define the relevant expertise and the exact job. 2. Context and source material: specify what I will paste and instruct the model to use it as the source of truth. 3. Method: give a short ordered process, including any required comparisons, calculations, criteria, or constraints. 4. Output specification: state the exact format, sections, word count or length, audience, tone, and exclusions. 5. Quality controls: tell the model what to verify, what uncertainty to flag, and when not to make assumptions. 6. Missing-information rule: ask up to 3 clarifying questions only when a required input is absent; otherwise proceed. After the finished prompt, provide: (a) a list of its fill-in placeholders, (b) one realistic example filled in for my task, and (c) two short test cases that would reveal a weak or ambiguous answer. Do not add invented facts, sources, business details, or policies. Self-check that every instruction supports [DESIRED OUTPUT] and that the prompt has no conflicting format, audience, or scope requirements.
Fill in
| Placeholder | What to enter | Example |
|---|---|---|
| [TASK TO COMPLETE] | Describe the job you need the AI to do in one or two sentences. | Turn customer interview notes into a product requirements brief. |
| [AI TOOL OR MODEL] | Enter the AI product or model where you will use the prompt. | Claude |
| [INTENDED USER] | Describe who will use the output and their relevant knowledge level. | A product manager preparing a planning meeting. |
| [DESIRED OUTPUT] | State the deliverable, its purpose, and the standard it must meet. | A one-page brief with user problems, evidence, assumptions, and ranked requirements. |
How to use
- Replace each bracketed field with the task, user, and deliverable you actually need.
- Copy the finished prompt the AI creates, then paste in the source material it needs to use.
- Check that its placeholders capture every fact the model would otherwise have to guess.
- Send this follow-up: "Revise the prompt for a first-time user and remove any instruction that does not affect the final deliverable."
Variations
Debug a prompt
Use this when an existing prompt gives inconsistent or generic answers.
Diagnose why this prompt fails: [CURRENT PROMPT]. I expected [EXPECTED OUTPUT], but I received [ACTUAL OUTPUT OR FAILURE]. Identify ambiguous instructions, missing inputs, contradictory constraints, and places where the model is forced to guess. Return a revised copy-ready prompt, followed by a short change log. Keep facts grounded in supplied material. Ask up to 3 questions only if the failure cannot be diagnosed from what I provided. Self-check that each change addresses a documented failure.
Create prompt variables
Use this when a team needs a repeatable prompt template.
Convert this one-off workflow into a reusable prompt template: [WORKFLOW]. It will be used by [TEAM OR ROLE] to create [DELIVERABLE]. Define 3 to 6 clearly named bracketed placeholders, explain the minimum acceptable input for each, and write one copy-ready prompt. Include a missing-input rule and output acceptance criteria. Do not create variables for information the model can infer safely. Self-check that every placeholder is necessary and that no required source is omitted.
Evaluate model output
Use this when you need a scoring rubric for AI-generated work.
Create an evaluation rubric for AI output from this task: [TASK]. The audience is [AUDIENCE], and acceptable output must satisfy [REQUIREMENTS]. Produce a 100-point scoring table with 5 to 7 criteria, observable pass/fail signals, and severity levels for factual errors, omissions, and formatting failures. Then write a short evaluator prompt that applies the rubric to an answer. Do not score style over correctness. Self-check that a reviewer could apply every criterion without guessing.
Tips
- Write the desired deliverable before choosing a model role; output rules usually matter more than a grand job title.
- Put source material boundaries in the prompt, especially when an answer must not draw on unsupported company facts or research.
- Use a small set of test inputs, including an incomplete case, before sharing a reusable prompt with a team.
- When results vary, inspect missing decision criteria before adding more background text.
FAQ
Should I tell the model to act as an expert?
Yes, but pair the role with concrete decisions, source boundaries, and output criteria. A role alone does not tell the model what good work looks like.
How many placeholders should a reusable prompt have?
Use only the inputs that materially change the result. Too many fields make a template harder to complete and less likely to be used correctly.
Why ask for test cases?
A test case exposes whether instructions survive real inputs, including incomplete or conflicting information.