AI for UX · Step 2 · 6 min read
Prompt patterns for designers
How to brief an AI tool the way you'd brief a colleague: a five-part brief structure, ten reusable patterns for design work, the mistakes that produce generic output, and how to keep a prompt library your team can trust.
Good prompting is mostly good briefing. You already know how to brief a colleague: who it's for, what you need, what the constraints are, how you'll judge it. A prompt is the same brief, written down. This note gives you a structure, ten patterns you can copy, and the habits that separate useful output from the internet's average answer.
What you'll be able to do after reading:
- Write a prompt with the five parts a model needs to do good design work
- Use ten patterns for common design tasks: options, critique, states, edge cases, copy, stakeholders
- Recognise the prompts that produce generic output, and fix them
- Start a small prompt library your team can reuse and improve
1. Why prompts matter so much
Remember from What AI is (and isn't): a model continues text in the most likely way. A short, vague prompt leaves almost everything open, so the model fills the gaps with the most average answer it can find. "Make a dashboard" gives you the dashboard the internet has seen most often.
Every detail you add closes a gap: who the users are, what they're trying to do, which components exist, what "good" means. The model doesn't get smarter; it gets less room to guess.
Prompting is one of three layers of context, each with its own place:
Stable rules every session reads: users, principles, tokens, components, "never" rules.
Typically: CLAUDE.md, AGENTS.md, DESIGN.mdWhat this feature must do: problem, users, flows, states, rules, acceptance criteria.
Typically: a spec in the wiki or repoThe request for this one iteration, inside the conversation with the tool.
Typically: "Show three options"Changes rarely → changes per feature → changes per iteration
Keep the prompt for the task. If you find yourself pasting the same paragraph into every prompt, it belongs in a context file.
2. The brief structure: RCTFC
Five parts, in any order. The examples use a fictional project: a booking app for a chain of community sports centres.
- RRole and context
Who is it for, in what situation?
- CConstraints
Components, accessibility, platform.
- TTask
Exactly what to produce.
- FFormat
What the answer should look like.
- CCriteria
How you will judge it.
A full prompt reads like a short brief:
Context: members of a community sports centre book courts on their phone, often one-handed,
between other tasks. Constraints: use only our components (Card, DateStrip, Button, Toast);
WCAG 2.2 AA; touch targets of at least 44 px. Task: give me three structurally different
layouts for choosing a time slot. Format: for each, a one-line idea, the components used and
one trade-off. Criteria: booking in under 20 seconds matters more than visual novelty.
You don't need all five every time. A quick rewrite of a button label needs a task and a constraint. A new screen needs all five.
3. Ten patterns for design work
| Pattern | Prompt to adapt | Why it works |
|---|---|---|
| Divergent options | "Give me 3 structurally different approaches, not variations of one idea." | Fights the pull towards one average answer |
| Critique me | "Act as a senior designer. Review this against these principles: … List the 5 biggest issues, most serious first." | A cheap first review before a person's |
| States sweep | "List every state this screen needs (empty, loading, error, no permission, very long content, offline) and what each should show." | Avoids happy-path-only design |
| Edge-case hunt | "What could go wrong for someone doing this task in a hurry, on a bad connection, or for the first time?" | Finds risks early |
| Explain your choices | "For each component you chose, explain why, referring to the brief." | Exposes invented logic |
| Stay in the system | "If a component you need doesn't exist in our library, say so instead of inventing one." | Prevents design-system drift |
| Spec from notes | "Turn these meeting notes into a feature spec using this template: …" | Faster first drafts of specs |
| Translate for stakeholders | "Summarise this design rationale for a finance manager in 5 bullets, focused on cost and risk." | Clearer communication |
| Content check | "Check this UI copy against these product terms: …; flag ambiguous or inconsistent labels." | Terminology quality |
| Localisation stress test | "Translate these labels into German and Finnish and flag any longer than 20 characters." | Finds layout breaks early |
4. Iterating: the conversation is the work
The first answer is a starting point. Good results come from a few rounds:
- 1React specifically
"Option 2, but price above time."
- 2Ask why
The reason reveals a misunderstanding or a good idea.
- 3One change at a time
So you can see what caused what.
- 4Start fresh when it drifts
A new chat with a better brief, or a context file.
steps 1 to 3 on every answer · step 4 only when the conversation drifts
5. Mistakes that produce generic output
| Mistake | Why it fails | Better |
|---|---|---|
| "Make it modern and clean" | An aesthetic without a purpose | Name the user, the task and the criteria |
| Everything in one giant prompt | Context, spec and task get mixed up; the model weighs them unevenly | Split: context file, spec, then a short task |
| Accepting the first answer | The first answer is the most average one | Ask for options, then direct |
| Asking AI to decide business rules | It will invent plausible rules | Ask it to list options; decide with your product team |
| Pasting real customer or employee data | It may break your organisation's rules and the law | Use realistic, anonymised examples. See Personal data in prompts |
| Trusting facts in the output | Numbers, quotes and sources may be invented | Check them. See Hallucination |
6. A prompt library your team can trust
A shared library turns one person's good prompt into everyone's starting point.
For each prompt, keep:
- Purpose: what task it's for, in one line
- The prompt, with placeholders like
<screen>or<user> - An example output, and what was good or wrong about it
- The tool and date it was tested with (prompts behave differently across tools and versions)
- An owner who updates it
Keep it where your team already works (a wiki page, a Figma file, the code repository) and review it every few months: retire prompts that no longer work.
7. Vocabulary
| Term | Meaning |
|---|---|
| Prompt | Everything you give the model for a task |
| System prompt | Hidden instructions the tool adds before your prompt |
| Few-shot prompting | Including a few examples of the output you want |
| Context files | Files a tool reads every time: rules and context you don't want to repeat |
| Temperature | How much the answers vary between attempts |
8. Try it (no code needed)
- Pick a small screen you know well.
- Prompt any approved assistant with only: "Design a better version of this screen." Save the answer.
- Prompt again with the full RCTFC brief (section 2). Save the answer.
- Run critique me on both, then compare. Which one could you actually use? What did the brief change?
Reflection questions: Which part of RCTFC do you usually leave out when you brief people, not just AI? Which of your team's rules should live in a context file instead of in every prompt?
9. If you lead a team
- Start a shared prompt library with five prompts that already work for your team, each with an owner.
- Agree that prompts containing real data are never shared or stored.
- Ask people to bring a before and after prompt to a team session: it's the fastest way to spread good habits.
Checklist
Learn more
- NN/g: Prompt to Design Interfaces: Why Vague Prompts Fail and How to Fix Them (Dec 2025)
- Anthropic: Prompt engineering overview
- W3C: WCAG 2.2, the accessibility baseline used in the examples
Books, courses and videos to go deeper
Related
Design specs for AI · Context files · AI design review process · Generating design with Claude Code · Figma AI · Claude Design · What AI is (and isn't)