Product design practice
Designing with AI
How I use AI across my product design workflow, and where I keep my hands on the wheel.
My stance
AI does the legwork. I do the judgement.
I treat AI as a fast, capable collaborator inside a user-centred process. It helps me make, research and synthesise. Deciding what is right for users stays with me.
Approach
Where AI fits in a user-centred process
| Stage | Where AI helps | What I keep |
|---|---|---|
| Discover | Pulling desk research, product docs and past findings into a starting brief | Choosing which evidence matters |
| Frame | Drafting problem statements and hypotheses in our house format | Agreeing the problem with the team |
| Ideate | Generating options and copy variations quickly | Picking the direction |
| Design | Building prototypes in Figma from our design system components and tokens | Every design and content decision |
| Test | Drafting test plans, then synthesising results with every theme linked to its source | Watching real users and reading the nuance |
| Measure | Turning results into clear stories for stakeholders | Deciding what happens next |
Context before prompts
Most of the value comes from what it knows before I ask it anything
Claude, connected to Figma, Maze, Confluence, Jira and GitHub.
Claude
Figma
Maze
Confluence
Jira
GitHub
Connected to my tools
Design files, research results, documentation, tickets and code, so it works from real files rather than my descriptions of them.
Taught our conventions
Design system tokens and components, brand voice and the house hypothesis format, captured as reusable skills.
Given standing rules
Preferences it follows every time, from punctuation style to never presenting a figure it can't source.
In practice
Three examples from real work
01 · Ideate
Exploring three directions before committing to one
What AI did
For the early termination charge journey, generated three distinct, clickable directions, each with its own design question to test.
What I did
Judged each against what members need to know before they leave, and chose which ideas to carry forward into the design.
02 · Design
Moving live journeys onto a new design system
Who did what
- What AI didRebuilt existing account screens in Figma on the new design system library, swapping old components for new ones and keeping layouts consistent from one journey to the next.
- What I didSet the rules it had to follow: components that match how the same thing appears elsewhere in the product, one format for related cards, and options hidden when a member can't use them.
03 · Design to code
Design and build in the same loop
What changed
On an internal tool, designs went straight into code, with generated components as the starting point instead of drawing every state by hand. I raise pull requests in GitHub for design changes myself, so they reach engineers ready to review rather than as a ticket to interpret.
Where judgement mattered
One generated pattern was rejected because it hid fault information from someone triaging a fault. Only a person holding that principle catches it.
Guardrails
What I never hand over
Real users stay real
AI never stands in for research participants. It helps me plan and synthesise, not replace the session.
Every number is traceable
Figures must come from a source, and anything unsourced gets challenged.
Sensitive data stays inside
Internal and customer data stays inside the business, whatever the tool.
I make the call
On design, on copy, and on what ships. The AI can suggest; it does not decide.
Failure modes
Where AI goes wrong, and what I do about it
01
Plausible, but not true
What it looks like
AI fills gaps with confident detail that sounds right but has no evidence behind it.
What I do now
Every claim is traced to real data before it reaches a customer. AI drafts; evidence decides.
02
Helpful, but off-goal
What it looks like
Left unchecked, it keeps adding helpful extras that dilute the one thing the user came to do.
What I do now
I brief with the goal and the success measure, not just the task, and judge output against both.
03
Generic, not ours
What it looks like
Without context, output drifts to generic patterns, values and tone instead of your own system.
What I do now
Our design system, brand voice and formats are loaded as skills, so the defaults are ours.
04
Persuasive, but overstated
What it looks like
Synthesis can smooth over nuance and make findings sound stronger than the evidence.
What I do now
Insights are rewritten in plain terms, with every stat and quote linked to its source.
Reflections
Six principles I work by
- 01Start with context, not prompts
- 02Work in small loops and review each one
- 03Keep a human checkpoint at every stage
- 04Verify anything that looks like a fact
- 05Codify what you repeat as rules and skills
- 06Be open about where AI was used
My role
How it has changed my role
Less
Production work
Resizing and re-laying out screens, first drafts of plans and copy.
More
Time on the problem
With users, with stakeholders, and on the decisions that actually matter.
Harder
Catching plausible but wrong
Output that looks finished can still be wrong for users. Spotting that is now a core skill.
In summary
Faster making. Same standards.
- AI is most useful when it knows your system, your users and your rules.
- The value is in the loop: generate, review, correct, repeat.
- Speed only counts if the work is still right for users.