Not an implementation detail. A design phase that blocks dispatch until the comps, the assets and the realistic data exist, so interface work is built against something real instead of improvised at the keyboard.
An agent asked to build a screen will not stop mid-task to become an illustrator. If the asset folder is empty when the work is dispatched, the screen ships as text on a background — every time, no matter how the brief is worded. Three interface rejections in a single day traced back to the same three causes:
Data work has no visual dependency, so it starts immediately and runs the whole time. Design runs beside it and ends at a hard stop: no interface code is dispatched until a human has seen the design and said yes.
Schema, repositories, business rules, domain logic, tests. Survives every design change intact. On the pilot product this track produced 400+ passing tests while the interface direction was rewritten three times.
Screenshot the real market and best-in-class references. Write down what separates funded-team execution from template execution in this category.
Image model produces visual-direction candidates and the character master sheet. Poses derive from that master by reference edit — never fresh prompts, which yield twelve subtly different characters.
Screen and state inventory, executable comps with real copy and realistic data, the asset pack, the fixtures. Committed as a baseline commit — implementation worktrees must fork from it or later, or they cannot see the art.
The whole app is browsable before a line of interface code exists. Criticism here costs a regeneration. The same criticism after implementation costs a rebuild.
planner refuses dispatch until packet passesWorkers build against the approved comp plus its layout specification, on top of the finished data track. The target is concrete, so "does it match" is answerable.
Every screen captured populated, empty, loading, error, and celebration where it applies. Screenshot and comp go to the critic side by side, with the scenario already mechanically verified.
Research across Stripe, Figma, Linear, Duolingo, Airbnb, Shopify and the Nielsen Norman Group's measured studies. Almost everything that separates funded-team work from template work turns out to be procedure with published effect sizes attached.
NN/g measured it: picking the best of four parallel designs is 56% better than the average. Merging the winner with the best parts of the rejects reaches 70%. Add one iteration and it is 152%. Generating one design and polishing it is the weakest available strategy.
Median usability gain from first to last version: 165%, about 38% per iteration, with measurable diminishing returns (45% on the first pass, 34% on the second). Below three total versions you are under the documented floor.
Heuristic evaluation runs 3 to 5 evaluators independently, each rating findings 0 to 4. A single rater's severity score is documented as unreliable; the mean of three is not. Three evaluators catch about 60% of all findings.
Expert designers agree with each other only weakly on absolute quality (Fleiss κ ≈ 0.30). Both humans and models are far more reliable at "A or B" than at a score out of ten, which is why UI benchmarks rank by tournament rather than rating.
The most-cited handoff failure is "the happy path is specified and nothing else." Loading, empty, error, success, plus hover, focus, active, disabled, plus validation and truncation. Everything else in a handoff is negotiable.
Stripe named its top 15 user journeys, walks each one quarterly with a mixed team, scores them red to green on a 5-point scale, and puts the result on a company scoreboard. Quality that is not measured does not survive growth.
Every role below is an agent lane, and every stop condition is mechanical. The operator picks the brand direction once. After that the loop runs itself, including the decision that the work is not good enough yet.
Screenshots the real market and best-in-class references, and builds the reference corpus that later percentile-scores every candidate. Metrics only mean something against a corpus.
One brand direction, the token set, the composition spec, the character master sheet. Poses derive from the master by reference edit so the character stays one character.
Three genuinely different executable comps per screen, assigned distinct theses so they cannot converge. This is finding 01: the cheapest quality available is generating more than one option.
Candidates compete head to head rather than being scored out of ten. The winner is then merged with the strongest elements of the losers, which is worth another 14 points on its own.
Each finding must carry expected standard → the gap → the fix plus a bounding box, rated 0 to 4. Severities are averaged across critics. Evidence-carrying findings mean a human can adjudicate any disagreement in seconds.
Any finding averaging severity 3 or higher sends the work back to the designers. Minimum three rounds regardless, because the measured floor is three versions. This is the loop that replaces a human saying "put another day into it."
iterate until no finding ≥ severity 3Contrast ratios, tap-target minimums, grid conformance, token conformance (no hardcoded values), the full states matrix, reflow at 320px, text-spacing overrides. Automated accessibility tooling catches about 57% of defects, so this layer is a floor and never the whole story.
Implementation screenshot against the approved comp, per state, per breakpoint. Deltas become tickets, not opinions.
A structural readiness check runs against this. Missing comp, missing fixture, missing declared anchor, asset that fails dimension or transparency validation — dispatch is refused. This is enforcement in the orchestrator, not another paragraph in a prompt.
Route, comp path, required states, fixture, and a visual anchor — character, data visualisation, photography, typographic composition, motion moment. A screen fails when it lacks its own declared anchor, not because it lacks a PNG.
The app emits a readiness marker naming the scenario and its record counts. Capture waits for it. This kills the classic failure where the seed succeeds, the app reads a different store, and the screenshot still shows zero rows.
Generating art in a parallel worker is useless if screen workers branched before it existed. Implementation worktrees are checked against the baseline commit.
Generated mockups invent text and impossible controls. They are excellent for direction and character art. The chosen direction is converted to executable comps with real copy and real data, and those become the reference.
A full studio pass for every colour tweak would make the system slow and expensive. The visual direction is chosen once per brand, then inherited automatically.
The pilot runs on three moments of one live product: populated home, core-action celebration, and a history screen with weeks of real data. It generalises only if first-pass operator acceptance measurably improves.