Process · 2026 edition
How the studio works
The first version of this document was written in 2017 for a federal science program: formal requirements, eight-point stories, PSDs handed to a front-end team, a two-sprint hypercare. Nine years later the steps are the same shape. The speed is not: the designer now ships the code, and AI helps with the coding, research and analysis. What has not changed: the work starts with people, and a person decides at the end of every step.
And the point of all of it is people. We want the software we touch to make someone’s day a little easier — the scientist and the dataset, the family and the clinic, the associate and the tablet. Technology is the means; the end is a fairer shot at getting the thing done. Then we check whether it worked.
Ground rules
Before any step, the standing law
-
Why. Speed without checkpoints just delivers the wrong thing sooner. A decision point costs an afternoon; an unexamined month costs the quarter — and the trust.
-
Why. Work that nobody signed gets relitigated by everybody. One accountable yes early beats ten drive-by nos late — and it tells the whole team whose call the next call is.
-
Why. Handoffs are where projects go to die. If design waits for research to be “done,” the findings are cold by the time anyone acts on them. Overlap keeps them warm.
-
Why. Requirements that start from the current API inherit the current API’s worldview. Start from the person, let the experience place demands on the plumbing — and nobody ends up shipping a screen shaped like a database.
-
Why. Drafting got cheap; judgment didn’t. A model can produce the option in minutes, but it can’t be accountable for a claim with a client’s name on it. The signature stays human, so the decision does too.
-
Why. The before-number is what makes the after-number mean anything. It’s the difference between “we redesigned it” and “we improved it 538%” — taste becomes evidence.
-
Why. People say true things when they trust the room. Leak the room once and every interview after it is theater — and research built on theater is worth less than no research.
-
Why. A mockup can’t feel latency, tab order, or a bad connection. The last hundred design decisions happen in code — so the designer has to be in the room when they’re made.
The steps
Nine steps, overlapping on purpose
Times are typical for a mid-size engagement with a responsive client — they stretch when recruiting is hard, and shrink when the brief is clear.
Measure the broken state
Know what is wrong, where, and how badly, before anyone draws anything: analytics, support tickets, a heuristic pass, a baseline survey, misclick data on the current screens. Out: a findings document written so it can be repeated.
Owner: designer + whoever owns the analytics · Time: 1–2 weeks (≈40–80 h) · AI drafts transcripts and coding; a person traces every figure to its source.
Frame it with the people who own it
Agree what problem is being solved, for whom, what the minimum useful release is, and how success is measured. Out: a one-page brief everyone who has to say yes has said yes to.
Owner: product owner, designer in the room · Time: 2–5 days of workshops, brief the next day (≈16–40 h).
Understand the people
Know who you're designing for from what they do, not what they say they do: interviews, ride-alongs, surveys in the org's own vocabulary. Out: a journey map — personas only when real clusters earn them.
Owner: designer · Time: 1–3 weeks; recruiting is the long pole (≈40–120 h) · Anything that could identify a participant is removed before it leaves the room.
Map the service and the system
See every screen, hand-off, and system a task crosses before deciding what to build. Out: flows including the failures, a systems inventory the roadmap is built on, a site map whose first release is boring and shippable.
Owner: designer + system owners · Time: 3–10 days (≈24–80 h) · Every inventory row confirmed with the person who owns that system.
Wireframe and decide
Settle structure, content, and behavior in low fidelity — real content, real labels, real error states, never lorem ipsum. Out: wireframes for every state plus a decision log: what was chosen, what it cost, what was left out.
Owner: designer · Time: 2–5 days per product area (≈16–40 h) · WCAG 2.2 AA checked now, not at the end.
Prototype and test
Prove the path works with the people who will use it — pass criteria set before anyone watches. Out: task success, misclicks, time on task, the verbatims that explain them, and the change list before build.
Owner: designer · Time: about a week (≈32–48 h) · The criteria never move to fit the result.
Visual design and the system behind it
Make it look like one thing on every screen, and make the next screen cheaper than this one. Out: tokens, a component library with variants, documentation a developer can build from.
Owner: designer · Time: 3–10 days for the system, hours per screen after (≈24–80 h) · Taste, hierarchy, and the accessibility sign-off are human calls.
Build
Ship the approved design as production front-end with the tests and accessibility checks in the code — HTML/CSS/JS, Salesforce Experience Cloud and Lightning, Node where it's a product.
Owner: designer + platform teams · Time: days to a few weeks (≈40–160 h) · Every diff read by a person and run on a real device with a screen reader.
Ship, watch, and measure again
Get it into people's hands, fix what the first week reveals, and run the step-0 measures again. Out: a before-and-after honest about which changes moved which number.
Owner: product owner; designer on call · Time: hypercare boxed at 2–4 weeks (≈16–40 h) — never open-ended.
End to end: about 250–700 hours of design and build for a mid-size engagement. The steps overlap on purpose, so the calendar is shorter than the sum — and every estimate above is an estimate, not a quote.
The human floor
What a person always checks
Where AI helps, with code, research and analysis, these six things never ship on its word alone.
Research claims
A person reads the source before a figure is written anywhere. Sample sizes are stated, not hidden.
Numbers
Every number on a page traces to a document, a dataset, or a published source. If it can’t, it isn’t on the page.
Accessibility
WCAG 2.2 AA is checked by a person with a screen reader — at wireframe and again in code. A tool’s pass is a start, not a sign-off.
Anything with a client’s name on it
Written or approved by a person. AI drafts are labelled as drafts until they are not.
Privacy
Participant data stays with the client. Models that train on inputs never see it.
Code
Every diff is reviewed by a person and run on a real device before it is committed.
Then and now
2017 → 2026, honestly
| Step | 2017 | 2026 | Faster by | Why — where AI carries it |
|---|---|---|---|---|
| Usability study | Several tickets over multiple sprints (≈4–6 wks) | 1–2 weeks, scheduling included | ≈3–4× (~70–75% less time) | AI transcribes sessions, codes the open text, tabulates the exports — a person still watches every session and owns every claim. |
| Wireframes, one product area | A 5–8 point story per sprint (≈2 wks) | 2–5 days | ≈3× (~65–75%) | AI generates variations to react against and drafts the microcopy; a person makes the call and writes down what it cost. |
| Visual comps | A 5-point story minimum; up to several sprints (2–6 wks) | 3–10 days for the system, hours per screen after | ≈4× (~75%) | AI produces the token scales, contrast checks, and tedious permutations; taste, hierarchy, and the accessibility sign-off stay human. |
| Design to live | PSDs and pixel values handed to a front-end team; quarters | The same person ships the front-end; days | ≈10× (90%+) | The handoff queue is gone. AI writes the first pass of components and tests; a person reviews every diff on a real device. |
| Hypercare | Two sprints after UAT | Boxed 2–4 weeks — then it ends | Same length — it just ends | AI summarizes tickets and logs daily so nothing lingers; a person closes the window on time. Discipline, not speed. |
The 2017 column is from the original 2017 document; the multiples are derived from the typical ranges shown, and engagement sizes vary — the direction doesn’t. The pattern in every row: AI carries the drafting, transcribing, and permutation weight; a person decides at the end of each step. That’s where the speed comes from — and why the quality survives it.
Next
See it applied
Every row on the work ledger ran through these steps. Twenty minutes on a call shows you what the document can't.