For a designer, prompt engineering means writing structured instructions that let an AI model produce a result you can use in a design workflow: an interview synthesis, flow variants, a documented component, a screen in code, a piece of microcopy. The method fits in seven blocks.
Why designers write bad prompts
Most designers use AI like a conversational search engine: a short question, a generic answer, disappointment. The problem is not the model, it is the brief. No designer would accept a one-line brief from a client. Neither should a model.
A good prompt looks like a good brief: it says who is speaking, for whom, in what context, with what objective, under what constraints, and what a good result looks like.
The seven-block structure
Here is the template I use for all my production prompts. It fits on one screen and reads like a specification.
ROLE Who the model should be (expertise, stance, level of rigor)
CONTEXT The product, the audience, the market, what exists, what has already been decided
OBJECTIVE The expected result, in one measurable sentence
CONSTRAINTS What is forbidden, required, limited (accessibility, brand, technical, legal)
STYLE The tone, the register, the references, what you absolutely do not want
OUTPUT The exact output format (structure, length, language, examples)
QUALITY The criteria that separate a good result from a plausible one
1. Role
The role sets the level of expertise and the stance. "Senior product designer and UX strategist, demanding, who challenges the brief" produces very different answers from "assistant." Also specify what the role does not do: "do not propose out-of-scope features."
2. Context
This is the most neglected block and the most important one. Product, audience, market, existing constraints, decisions already made, the team's vocabulary. The more precise the context, the less the model invents. An existing design system, brand guidelines, personas: paste them in.
3. Objective
One sentence, one verifiable result. "A clear, appealing and measurable onboarding that gets the user to first value in under three minutes" beats "improve the onboarding."
4. Constraints
WCAG AA accessibility, mobile-first, tokens v3, no new colors, no modals, French language, maximum label length. Constraints are what make the result buildable.
5. Style
The editorial or visual register: "editorial, high contrast, restrained motion," "direct tone, short sentences, no jargon." Add the no-gos: "never use em dashes, never use emoji, never use superlatives."
6. Output
The exact format: a table with specific columns, a flow in numbered steps, a React component with typed props, microcopy of 40 characters maximum. Give an example of the expected output whenever you can.
7. Quality
The criteria to judge by: "hierarchy readable in under three seconds," "one primary action per screen," "every error message says what to do." Ask the model to check its own output against these criteria before answering.
Three complete prompts
User interview synthesis
ROLE Senior UX researcher, rigorous, never generalizes beyond the data
CONTEXT B2B SaaS for managing workwear allocations. 8 interviews of 45 min with
HR managers. Transcripts below. Starting hypothesis: tracking returns is
the main friction point.
OBJECTIVE Identify the 5 major frictions, each backed by at least 2 quotes
CONSTRAINTS Do not invent quotes. Flag when the starting hypothesis is not confirmed.
Separate what is said from what is interpreted.
STYLE Factual, concise, no superlatives
OUTPUT Table: friction · frequency (n/8) · quotes · interpretation · design direction
QUALITY Every row must be verifiable in the transcripts. End with what the
data does not allow you to conclude.
Generating a component's missing states
ROLE Design system lead, obsessed with consistency
CONTEXT Design system v3, tokens attached. Input component: default, focus and
error states exist. Naming convention: state/variant.
OBJECTIVE Complete the component's missing states
CONSTRAINTS Reuse existing tokens only. No new tokens. AA contrast.
STYLE Same density and same radius as the existing components
OUTPUT For each missing state: name, trigger, appearance (tokens used), accessibility
rule, sample microcopy if relevant
QUALITY List the states you did NOT add and why
Error microcopy
ROLE Senior UX writer, native French speaker, direct tone
CONTEXT Checkout flow for a subscription. Premium brand, formal address ("vous").
List of technical error codes below.
OBJECTIVE One message per error, understandable by a non-technical person
CONSTRAINTS 90 characters maximum. Always say what to do next. Never any jargon,
never an exclamation mark, never an em dash.
STYLE Calm, precise, respectful
OUTPUT Table: code · message · suggested action · short variant (50 chars)
QUALITY Reread each message and ask: does a stressed user get it in one read?
Version, test, share
A production prompt is treated like a component:
- Version it: one file per prompt, a version number, a changelog.
- Test it: three to five real input cases, an expected output for each, and a review at every change.
- Share it: a single location, examples, an owner. A prompt nobody maintains drifts.
This is what I call prompt design: designing prompts as interfaces for the team. See the prompt engineering for design page.
What the prompt will not replace
Understanding the problem, making trade-offs between conflicting constraints, owning the result in front of a client or a user. The model amplifies an intent. If the intent is vague, the result will be too, only more confident.
To go further: Generative AI in a product design workflow.

