The key idea
Start with the deliverable, not a role. ‘Act as a marketer’ is optional; it cannot replace sources or acceptance criteria.
Try it with LumaDesk
Six things to specify
Goal
Name the work to produce.
Audience and context
Explain who will read it and where.
Approved sources
Separate evidence from instructions.
Constraints
Set limits and prohibited claims.
Output format
Specify sections and lengths.
Review criteria
Define what passes and fails.
Try it with LumaDesk
Fictional exercise; the reference below was written for this guide.
A — Approved: shared editorial calendar, approval tracking; French and English. No pricing or performance figures.
B — Older, unapproved draft: German support; 30% time saving.
Reusable prompt
Using A and B above, write homepage copy for LumaDesk aimed at independent agencies. Only A authorizes product claims; use B solely to flag conflicts. Produce three headlines of at most 12 words each, one 60–80-word paragraph and a list of missing facts. Do not invent pricing, German support or time savings. Mark missing information unknown. Before returning the copy, check every claim against A and verify the lengths.
Reference outcome
The copy can mention the calendar, approvals and two supported languages. Pricing and measured results remain unknown; B’s conflicting claims are flagged, not adopted.
Pass or fail?
- Every product claim is supported by A.
- Missing facts are marked unknown.
- Headlines and paragraph respect the limits.
- B’s conflicting claims are not silently merged.
Download the complete exercise pack
Improve one thing at a time
Fix one recurring issue, then replay the same three cases: complete sources, missing pricing and a conflicting old draft.
Further reading
Prompt engineering for marketing teams
Create reusable instructions that make the task, evidence and review criteria clear.
A useful prompt specifies the objective, audience, source material, constraints and output format. It does not replace a clear brief or reliable evidence.
Our method
Provide approved context and ask the model to identify missing information before drafting. Use examples to clarify tone. Separate research, drafting and checking into reviewable stages. Save the prompt with the input conditions and version that produced the result.
How to assess progress
Test the same instructions on normal and difficult inputs. Compare factual accuracy, usefulness and correction effort. Avoid judging a prompt from a single impressive response.
Decision
Decide which part of the output must become more reliable: selecting evidence, following a format, matching the audience or recognising missing information. Change the prompt to address that specific failure. Adding instructions without diagnosing the problem makes later improvements difficult to attribute or maintain.
Exercise
Choose five briefs and reserve two more for a final check. Run the original prompt, label each correction by failure type and change one instruction at a time. Include an incomplete brief and contradictory source material. Require the output to mark unresolved facts instead of supplying plausible details to satisfy the requested format.
Deliverable
Deliver a prompt card containing purpose, required inputs, instructions, output structure, examples and known limitations. Add a small change log connecting each revision to an observed failure and its retest. Store the reference outputs beside the card so a colleague can judge whether future changes improve the same task.
Pitfall
An overly specific example can cause the model to copy its content into unrelated assignments. Test different products and audiences before accepting the prompt. A successful run on the brief used to write the instructions is development evidence, not proof of general reliability.
Official sources
OpenAI · Prompt engineeringAnthropic · Prompt engineeringOpenAI · Evaluation best practicesWant this workflow built for your team?
Help your team use AI consistently in real work. Connect training, review responsibilities and practical documentation.
Discuss my AI workflow