Designing a digital experience with AI starts with context.
You don't open the laptop and start prompting. You build context first — the proper files, born from real research. This is how I use the Double Diamond to gather that context, stay on track, and only then hand the folder to an AI agent to ideate and prototype.
A guided walkthroughWorked example: the Brink Activation Brief~1 hour
01 — The premise
An agent is only as good as the context you give it
Everyone is learning the same lesson at the same time: the bottleneck in designing with AI isn't the model, and it isn't your prompt. It's context. A capable agent handed a vague sentence produces vague, generic work. The same agent handed a folder full of real research — interviews, a sharp problem statement, measurable goals — produces work that is specific, grounded, and yours.
So the question becomes very practical: what is the context, and how do I produce it? The answer is the unglamorous part nobody can skip — you do the research. You talk to people, you synthesise what you heard, you name the real problem before you reach for a solution. That work doesn't just inform the design; it becomes the files the agent reads.
I use the Double Diamond to do this consistently. It keeps me honest about staying in the problem before I jump to the solution, and it gives the whole process a shape I can repeat on every project. By the time I reach the agent, I'm not starting a conversation — I'm handing over a briefed teammate.
Context is a deliverable. The research you do isn't a phase you leave behind — it's the payload the agent works from.
02 — The method
The story of the Double Diamond
In 2005 the British Design Council went looking for a simple way to describe how good design teams actually work. They studied the process across very different organisations and drew it as two diamonds in a row. It has stuck for twenty years because it captures one stubborn truth: good design separates understanding the problem from solving it.
Each diamond is a breath — out, then in. The widening half is divergent thinking: explore widely, gather more than you need, resist conclusions. The narrowing half is convergent thinking: synthesise, decide, commit. You do this twice.
First diamond — the problem space
Discover then Define. Are we even solving the right problem? You don't earn a solution until you've earned a sharp problem statement.
Second diamond — the solution space
Develop then Deliver. Now that the problem is sharp, explore many solutions, then converge on one worth shipping.
The pinch in the very middle is the most important point in the whole model. It's the moment the problem becomes a single, defensible sentence. Everything before it is about earning that sentence; everything after it is about answering it. In our world, that midpoint is also the moment the folder is rich enough to hand to an agent.
Watch the diamond fill
As you scroll, the diamond in the bar above colours in — one quarter per phase. We'll fill the first half (Discover), complete the first diamond (Define), cross into the second (Develop), and complete it (Deliver). Four fills, one full process.
Diamond 1 · First halfdivergent — go wide
03 — Discover
Discover: gather requirements, collect research
The first half of the first diamond is research, and it is divergent on purpose. You are trying to understand the people you're designing for in their own words — separating the real blockers from the ones you assumed. You gather more than you'll use, because you don't yet know which thread is load-bearing.
Six ways to gather the evidence
There is no single "right" method — you choose the mix that fits the question and the people. These six cover most generative discovery:
Interviews
1-to-1, semi-structured. The depth method — the "why" behind the behaviour.
Questionnaires / surveys
Scale and breadth. Confirm whether a pattern is common or anecdotal.
Diary studies
Self-reported moments over days or weeks. Catches what people forget to mention.
Focus groups
A handful of people in conversation. Surfaces shared language and disagreement.
Contextual inquiry
Watch people use the real thing, thinking aloud. Behaviour, not opinion.
Analytics / desk research
The funnel numbers and prior studies. Tells you where to point the qualitative work.
What the research turns into
Raw evidence isn't context yet — it's noise until you shape it. Discovery's job is to convert conversations into a small set of durable artifacts. These five are the backbone of the folder:
Empathy map
What the user says, thinks, does and feels in one charged moment.
Pain points
The problems, ranked by how badly they hurt and how often they're hit.
User personas
The archetypes you design for — and the anti-persona you don't.
User stories
Capabilities as "As [user], I want… so that…", traced to evidence.
User journey map
The current experience end to end, with the emotional curve that shows exactly where it breaks.
How Discover looked on Brink+
Brink is a Dutch two-sided talent marketplace. Only about 1 in 4 professionals who sign up ever become "assignment-ready." Discovery meant six 45-minute interviews with a think-aloud walkthrough, deliberately sampled across the spectrum — people who finished, people who abandoned, and one who left for a competitor — so the synthesis wasn't rigged toward a conclusion.
Five usable transcripts produced seven insights, each traceable to at least two participants. The clearest voices:
"I don't want to author a profile from nothing. I want to correct one — and then get back to actual work." — Sanne, experienced interim PM
"I put a rate in, deleted it, put a lower one, felt like I was underselling, deleted it again." — Mehmet, newly independent
Those conversations became the empathy map, nine ranked pain points, two personas (plus a veteran anti-persona), user stories, and a current-state journey map whose low point is unmistakable: the rate field, faced alone, at the emotional bottom of the journey.
Diamond 1 · Second halfconvergent — narrow down
04 — Define
Define: name the problem before you touch a solution
The second half of the first diamond is where the diamond fills completely. You take everything discovery surfaced and squeeze it — past symptoms, past your first ideas — until what's left is a single, sharp articulation of what you're solving and for whom. This is the discipline most people skip, and it's the one that makes everything downstream cheap instead of expensive.
Three artifacts do the converging:
Problem statement
The convergence point. One paragraph, solution-free, that names who is stuck, where, and why — ideally with a root-cause chain so you solve the cause, not the symptom.
Hypothesis statement
The bridge from "what's wrong" to "what we believe will fix it." Written so it can fail: it names the change, the expected measurable movement, and the reason. If you can't be proven wrong, it isn't a hypothesis.
Value proposition
The promise to the user, in their language — the thing the solution must deliver. It turns the problem into a reason to care.
Solve the cause, not the symptom. Brink's real problem wasn't a bad rate field — it was a funnel designed around profile-completion instead of a real match.
Brink's problem, hypothesis & promise+
Problem statement. "New independent professionals abandon Brink's activation flow in the messy middle… because the experience front-loads duplicated effort, forces a high-stakes rate decision with no reference, and withholds any visible reward until the very end. The path from interested to active is built backwards — effort first, reward last."
Hypothesis (with the riskiest part flagged). If we draft the profile from an imported CV/LinkedIn, give transparent rate guidance, and surface a real match preview at ~halfway readiness, then activation rises from 34% to ≥50% and time-to-ready falls from 9 days to under a day. The riskiest assumption — test it first — is the early match preview: shown badly, it could read as fake and erode trust.
Value proposition. "Brink gets independent professionals from sign-up to a real, relevant match in minutes — not days — by turning profile-building into a guided draft you edit instead of a form you author… and proving there's work for you before asking for the boring stuff."
The midpoint of the Double Diamond is a single sentence you can defend. On Brink, that sentence is: the funnel is built backwards.
Hand the folder to the agent: prepare the environment to ideate
We've reached the pinch. The first diamond is fully coloured in, and the folder now holds everything a teammate would need to start: research, personas, a journey map, a sharp problem, a tested-able hypothesis, measurable goals. This is the exact moment the work stops being a document and starts being context for an AI agent.
Before ideating, I set up the environment so the agent can actually use what I've built:
① Drop in the project docs
The research folder itself — every artifact from the first diamond — sits in the repo where the agent can read it.
② Add skills
Reusable instructions for recurring jobs (writing in your voice, building a prototype, running a critique) so the agent works the way you work.
③ Craft CLAUDE.md with /init
/init writes the project's brain: what this is, where things live, the rules to follow. The agent reads it first, every session.
④ Hook up the MCPs
Connect the tools — Figma, the design system, a data source — so the agent can act, not just talk.
# in the project folder/init# generate CLAUDE.md — the project's brain/context# confirm the research docs are loaded# skills in .claude/skills/ · MCP servers connected · ready to ideate
The pointPrompting well matters far less than preparing well. A briefed agent with the folder, the skills, the brain file and the tools will out-design a clever prompt with none of that — every time.
Diamond 2 · First halfdivergent — explore solutions
06 — Develop
Develop: ideate and prototype, with the agent
The second diamond opens and we go wide again — but this time in the solution space, with the problem already sharp behind us. The diamond fills to three-quarters here. This is where the agent earns its place, because it can generate and revise design directions faster than any of us can sketch them.
Set the target, then study the field
First a product goals statement — one goal, one population, one timeframe, with an explicit "without" clause that forbids the cheap win. Then a competitive audit: not a feature checklist, but a real look at how others handle the visual, the interactions, the content, and what each company actually does well or badly. The audit's payoff is finding the open lane — the thing nobody has solved.
The open lane. Nobody combined early reward + transparent AI for complex, multi-sector work. That gap is the design opportunity.
"How might we do our own version?"
The competitive gap and the problem statement reframe into How-Might-We questions — open enough to invite many ideas, narrow enough to stay on the real problem. Then the agent does what it's good at: generate a rough first draft, fast. A draft you can react to is worth ten abstract discussions.
From that draft you iron out the user flow and outline the information architecture. This part is best done with a few people in the room — fresh eyes keep ideas growing instead of calcifying around the first one. The agent is the tireless sketch artist; the humans keep the judgement. With a clear direction (or two) and a flow you trust, you slide straight from ideation into prototyping.
Brink's ideation: from HMW to three concepts+
The primary How-Might-We: "How might we let professionals feel that work exists for them before they've finished setting up?"
That, with the rate and authoring reframes, produced three connected concepts to prototype as one flow:
Brink Intro — import-first profile building; you correct a draft instead of authoring from a blank page (express lane for veterans, guided lane for newcomers).
Match Preview ★ — a real, relevant assignment surfaced at ~halfway readiness. Highest leverage and riskiest — so it's the first thing to prototype and test.
Rate Copilot — the empty rate field replaced with a transparent, editable market range the user always confirms.
Scored on Impact × Confidence × Effort, governed by one guardrail running across all of them: the AI is transparent, editable, and never auto-published.
Diamond 2 · Second halfconvergent — converge & ship
07 — Deliver
Deliver: prototype, test, refine, hand off
The final quarter fills and the process completes. Develop gave us directions; Deliver narrows them to one thing, real enough to put in front of people and solid enough to hand to engineering. With the agent in the loop, this convergence is faster and tighter than it has ever been.
Prototype to a shareable versionBuild it up — with the agent — to something you'd happily put in front of the client. Lead with the riskiest assumption first.
Get feedback from real usersPut the prototype in front of the people from your research. Watch where it holds and where it breaks.
Feed the feedback back into the agents, and repeatEach round of feedback becomes new context. Iterate until the prototype is ready for delivery — measured against the goals you set, not vibes.
Hand the coded prototype to engineeringStore it in GitHub and give the engineers access to a working, coded prototype — not a static mockup they have to reinterpret.
Deliver is a loop, then a handoff. Refine against evidence until ready, then give engineering something coded — not something to guess at.
Measured against the goal, not a feeling
Because Define gave us numbers, Deliver can be honest. On Brink the contract was explicit: lift 7-day activation from 24% to 40%, profile-to-ready from 34% to ≥50%, time-to-ready from 9 days to under one — without degrading profile quality or trust. A prototype either moves those or it doesn't.
The contract. The metrics set in Define are what Deliver is judged against.
08 — And again
One full diamond — and a loop you can now run fast
Both diamonds are coloured in. We went wide to discover, narrowed to a sharp problem, went wide again to develop, and narrowed to something shipped. That's the whole shape — and the real unlock is that it's repeatable. Real design isn't one clean pass; it's this loop, run again and again as you learn.
What's changed is the speed. The discipline is the same one designers have used for twenty years — but with the research banked as context, a briefed agent, and the tools wired in, each turn of the loop is a fraction of what it used to cost. You keep the judgement. The agent takes the grind.
The skill isn't prompting. It's preparing the context — and the Double Diamond is how you prepare it well.
Read nextEverything the first diamond produced for Brink is written up in the companion User Research Document — the actual context you'd hand to the agent.
