What happens between the moment requirements land and the moment an engineer starts
| 01 | Intake: the first 48 hours |
| 02 | The workspace: the skill library |
| 03 | Mapping: six directions |
| 04 | Handoff: what developers get |
| System | Where | Role |
|---|---|---|
| Edison | GE Healthcare · 40+ clinical products | Core contributor · Red Dot 2020 |
| FIBER | Blue Cross Blue Shield · three teams | Built from nothing to 54 components |
| TUCM | TransUnion · global, multi-region | System and visual identity |
| Pathfinder | Avant · regulated lending | Sole design-systems contributor |
| Enterprise | Verizon · iOS, Android, web | Post-rebrand audit and token layer |
The first week produces no screens on purpose. It produces the context everything after it is generated and judged against.
The problem in the client’s own words. The metric it has to move. The users and the job they’re hiring the product for. Hard constraints from engineering. Assumptions, ranked. Open questions, owned and dated.
It is the context every later step reads from, and the thing re-read when scope drifts. A model amplifies whatever context it is given, so a vague brief buys confident, plausible, wrong output at speed.
What the skill does and when to use it. At rest, this is the only part in context. It is how the model knows the capability exists at all.
The steps, the rubric, the output format. Loaded only when a task matches, so a library of ten costs almost nothing until one is needed.
Templates, checklists, schemas and scripts it opens on demand. Rubrics and report formats live here, versioned like code, not retyped.
| Skill | What it does | What ships with it |
|---|---|---|
| brief-intake | Turns a raw brief into the structured project file | Intake template, question bank |
| heuristic-audit | Rates a live product against heuristics and WCAG 2.2 | Rubric, severity scale, report format |
| ds-inventory | Inventories components, extracts tokens, reports drift | Token schema, coverage matrix |
| research-synthesis | Transcripts into insights with counts and implications | Evidence rules, insight template |
| Skill | What it does | What ships with it |
|---|---|---|
| flow-map | Flows, plus the full state matrix for every screen | State checklist, notation |
| copy-standards | Errors, empty states and labels in the product’s voice | Voice guide, message patterns |
| a11y-spec | Focus order, roles, labels and contrast pairs per screen | WCAG 2.2 checklist |
| handoff-spec | Annotated spec, token export, state matrix, QA list | Spec template, QA checklist |
| critique | Argues against the work before the team sees it | Rubric, common-failure list |
The question is coverage and consistency, so the work starts with artifacts: the code, the library, every screen that ships. It is a counting exercise before it is a taste one: what exists, what’s duplicated, what has drifted, what has no owner. Verizon’s post-rebrand system started exactly here.
The question is whether the thing does the right job, so the work starts with people: who uses it, what they are trying to finish, where they give up. Artifacts come second. A perfectly consistent product can still be the wrong product. SurePeople’s PRISM platform started here.
| Finding | Count | Cost | Fix |
|---|---|---|---|
| Greys not on the ramp | 23 | Re-theming is impossible | Map to four steps |
| Button variants in code | 9 | Every screen re-decides | Collapse to three |
| Components with no focus state | 14 | Keyboard users stall | Fix the primitive once |
| Screens with detached instances | 31% | Fixes stop propagating | Re-link the top ten flows |
| Spacing values in use | 17 | Rhythm reads accidental | Snap to the 8-point scale |
Every screen resolves to a number someone is accountable for. If the path from pixel to P&L can’t be traced, the work is decoration.
Every element resolves to a token and a component. Anything that can’t is either a gap in the system or an indulgence.
The same pattern wherever it already appears. Most problems have been solved once in this product already. Adopt, don’t invent a second one.
What the API actually returns: slowly, partially, stale, or not at all. The data model is a design input, not an implementation detail.
Every state the screen will pass through, and what it looks like at ten times the content and three years of accumulated cruft.
Keyboard, screen reader, small screen, long strings, other languages, no network. The edges are where the product loses people quietly.
54
Production components in FIBER, adopted by three teams, with a coded library that matched the design source in naming, structure and behavior. Parity is the whole point. A component that doesn’t match its code is a suggestion.
| Artifact | What’s in it | What it prevents |
|---|---|---|
| Token export | Color, type, spacing, radius, elevation | Translation and eyeballing |
| State matrix | Default, hover, focus, active, disabled, loading, error, empty | States invented in code |
| Annotated flows | Behavior written as rules, not redlines | Breakage on real content |
| Accessibility spec | Focus order, roles, labels, contrast ratios, 44px targets | A final-week audit panic |
| Design QA list | Checked on staging before release | Handoff ending at delivery |
The tooling compresses the loop. It doesn’t decide what the loop is for.
Given a brief and an hour: the intake file, the skills that fire, and the first map, on your product
mikelrosenthal@me.com mikelrosenthal.com