When a prompt disappoints, diagnose the missing instruction instead of adding more impressive language. A useful diagnostic has five parts: role, context, task, constraints and format. Use only the parts the task needs, then check the result separately.
This is an editorial template, not a vendor standard. OpenAI’s prompting guide discusses instructions, context and examples; Anthropic’s overview starts with success criteria and a way to test them. Neither makes these five labels mandatory.
The five parts
| Part | Question it answers | Useful example |
|---|---|---|
| Role or stance | What approach should the assistant take? | “Edit for a reader unfamiliar with the project.” |
| Context | What source facts and audience matter? | “The event moved from Tuesday to Thursday; the venue is unchanged.” |
| Task | What should it do? | “Draft a notification.” |
| Constraints | What must be preserved or excluded? | “Do not invent a reason, time or apology.” |
| Format | What shape serves the reader? | “One paragraph under 70 words.” |
A role is optional. “Careful editor” is shorthand for behavior, not a professional credential. Concrete instructions such as “quote the sentence before proposing a change” are easier to assess than a biography claiming twenty years of expertise.
Check the input before improving the prompt
Practice with fictional material. For real work, establish permission for the exact account, feature, purpose and data before pasting context or attaching a file. A relevant detail can still be confidential. ChatGPT data controls describe product handling; they do not authorize client or workplace material.
Give the model the smallest sufficient source. Separate supplied facts from suggestions and unknowns. If the venue is undecided, say so rather than expecting a model to infer that from its absence.
A complete prompt and a checked result
This is an authored teaching example with fictional facts and output, not an observed product run.
Role: Help me write a clear client reply.
Context: Our agreed deliverable is a three-page website. The client has asked
for two additional pages. I am willing to discuss a separate quote.
Task: Draft one reply explaining the boundary and next step.
Constraints: Under 80 words. Friendly and direct. Do not invent a fee,
delivery date, contract clause or promise that I have capacity.
Format: One paragraph, no greeting or sign-off.
A checked draft:
“The current deliverable covers three pages, so the two additional pages would need a separate quote. If you would like to explore that, send me what each page needs to do and I can review the request before we agree any change.”
The draft preserves the three-versus-two distinction, offers discussion of a quote, and promises no price or delivery slot. The last sentence proposes a request for information; keep it only if that is the next step you intend. “I can add them next week for €400” would fail two explicit constraints.
Validation is the sixth step, not a magic instruction
A prompt can request missing-field labels or an assumptions list. You still need to compare the output with the input. NIST’s generative-AI risk profile describes confident false content, which can appear even in a neatly formatted response.
For the example, check the scope, proposed action and absence of new commitments. For a summary, check claims and exceptions against the source. For a calculation, recalculate independently. Do not let the assistant’s statement that it “followed every instruction” replace those checks.
Repair the failed part
If the reply invents a deadline, repair the source boundary or the constraint. If it is too long, change the format or length requirement. If it answers a different question, make the task verb explicit. Missing data may require a question; a different model or tool may be necessary for a capability gap.
OpenAI’s ChatGPT prompting advice recommends iterative refinement. Use that as a practical loop: identify one failure, change what addresses it, then compare the next output. Save the task-contract worksheet once a recurring prompt has passed ordinary, missing-input and conflicting-input cases.



