Authoring Hi Fi

authoring-hi-fi — SKILL.md

Variant: standard · When to use: producing (or amending) a high-fidelity UI design as runnable, rendered code from the upstream wireframes + design-system — a build seed an engineer can grow into production, gated on a human visual check.

Overview

This skill is the how-to of producing strong, comprehensive high-fidelity UI design as code — the judgment a producer applies, not the section list. The artifact is code that renders to pixels: standalone HTML+Tailwind by default, React+shadcn when the build stack is React (so the seed sits close to the build target and minimizes drift). It assumes a template tool (section structure), a headless browser (agent-browser/Playwright, to render + screenshot), the product's design-system + shadcn (the token + component source of truth), and a research capability. The producer is handed the upstream wireframes (the screens + states + structure) and the design-system (the tokens + component contracts) and realizes every wireframe-named screen — never a blank page, never re-deciding structure.

The signature mechanism is a loop, not a single write: generate → render headlessly → screenshot (per viewport × state) → vision self-review against the DS + wireframe + the bar → refine → repeat to the bar → human visual gate. This requires a vision-capable runtime; a text-only run is blind codegen (degraded — say so, never fake a render).

Two boundaries keep this in lane: hi-fi owns final visuals realized as pixels + the numeric pixel-WCAG judgment (the wireframe only reserved these); it consumes the design-system's tokens (never invents/grades them); it renders the given wireframe structure (never re-decides layout/screens/nav); and it is a build SEED — intentionally thin (no tests, mock data, CDN fine, single-file fine) — not final production code (the build phase rebuilds the data layer, routing, state, tests, perf).

When to activate

• Authoring a new hi-fi design as code from the upstream wireframes + design-system.
• Realizing the key screens (+ their states) a wireframes doc specifies, to rendered pixels.
• Amending an existing hi-fi when a token, screen, shared layout, or referenced component changes (see Step 8).
• Filling a hi-fi template with researched, rendered, decision-complete screens.

Do NOT activate when:

• Reviewing or grading a finished hi-fi → use reviewing-hi-fi.
• Producing the screen structure (layout regions, component selection, states-as-structure) → that is the upstream wireframes.
• Authoring/grading the token system / component catalog → that is the design-system; consume it here.
• Re-deciding which screens/transitions exist → that is user-flows.
• Writing final production code (tests, real backend, routing) → that is the build phase; this is a seed.

Inputs

Read every document the plan hands you — your depends_on set (the typical upstreams: the wireframes that name + structure the screens, and the design-system that supplies tokens + component contracts). The named upstreams are guidance, not a precondition: be self-contained — produce from whatever you receive; when an expected upstream is absent, proceed on what you have and surface the gap as an assumption (Step 7), never fabricate. Confirm the runtime can render + you can see the render (a vision-capable runtime + a headless browser); if it cannot, say so and treat the run as degraded — do not present an unverified render as reviewed.

Capability context (when provided): If a capability_record (a record from capability-map.yaml product_capabilities) is injected by the caller, read it before Step 1. It defines your scope boundary: owns = entities you cover; refs = entities you reference but do not own; publishes/consumes = events you surface; entry_points/exit_points = how users arrive and leave; has_ui/has_api/has_persistence = which surfaces apply. When present, treat it as a hard constraint — do not stray outside the boundary it defines.

Workflow

Step 1: Take the structure from the template tool — don't invent an outline

Get the section structure from your hi-fi template tool (comprehensive variant). Do not restate a section list here; this skill supplies the method that fills those sections well. If none is available, obtain/forge a comprehensive hi-fi structure, then proceed.

If a capability_record is present: apply design-system tokens to the per-capability wireframes only — do NOT redesign the shell or other capabilities. If scope=system: apply to the system-wireframes (shell) only. Respect the same boundary as the wireframes document you are polishing.

Step 2: Derive the screen set from the upstream wireframes

The hi-fi depends_on the wireframes. Every wireframe-named screen, and every state the flow/wireframe visits, gets a rendered hi-fi screen. Build the screen-inventory (each screen + its states + target viewports → its section) so coverage is auditable — this is also the amend-ripple surface (Step 8). A wireframe-named screen with no defined content is a gap — record it (Step 7), never invent it. Render the given structure; do not add screens/regions/nav the wireframe didn't name (an over-reach).

Step 3: Set up the token → code map (consume the design-system; never invent)

Map the design-system / DTCG tokens into the code's theme layer once, then bind everything to it: semantic CSS custom properties (--background, --primary, … with a .dark override), consumed by Tailwind utilities (hsl(var(--token)) in the config, or the v4 @theme block) and the shadcn CSS-variable theme on a React stack. Every visual value is token-backed — no raw #hex / 12px / bg-blue-500 that bypasses a token. A token/component the DS lacks is a surfaced gap (→ amend the DS), never inlined. (Depth: references/render-loop-and-tokens.md.)

Step 4: Realize the visual language + polish to pixels

Turn the wireframe's structure + the DS tokens into a polished, real screen:

• Visual execution — token-backed color, real typography (type scale/weight/line-height/measure), the 8pt spacing cadence realized, elevation/shadow, radius, a real icon set, placeholder-safe imagery; the wireframe's structural hierarchy now expressed in real contrast/scale/weight/color/position.
• Objective polish (checkable) — aligned to the grid, consistent spacing scale, optical balance, no clipped/overlapping/orphaned elements, on-token values, consistent treatment across sibling screens.
• The five named aesthetic heuristics (the program's polish bar — judged, but each named + articulable; subjective out-of-set taste is NOT a requirement): balance (visual-weight distribution), whitespace & restraint (intentional negative space; not cramped, not over-decorated), spacing rhythm (a consistent, pleasing cadence), focal clarity (one clear emphasis the eye lands on), aesthetic cohesion (a "designed", intentional feel — not assembled-from-parts). (Depth + "what good looks like" per heuristic: references/visual-a11y-and-amend.md.)

Step 5: Render real content + every state, then run the render→vision-review→refine loop

• Real content (no lorem) + realistic + edge data (long strings, overflow, empty, pluralization, i18n length) — lorem hides layout/overflow/tone/a11y problems.
• Render every state the flow visits, as actual pixels: empty (reason + guide-to-action), loading = a skeleton mirroring the populated layout (not a bare spinner), populated, error (plain-language cause + recovery), success (what changed + next step), partial.
• Run the loop: generate the code → render it headlessly (agent-browser) → screenshot per viewport × state → vision-review the screenshot against the DS + wireframe + the Step-6 bar → refine → repeat. The loop is bounded (§9): cap the iterations; if a screen still can't reach the bar (e.g. a missing DS token caps contrast), stop and surface the blocker as an assumption/gap — do not spin. Present the refined screens + their screenshots, never first-shot codegen.

Step 6: Self-check against the bar before handing off

Confirm all hold (this is the bar reviewing-hi-fi asserts — same list, single-sourced, each proportional to the screen's archetype):

1. Full coverage — every wireframe-named screen + state is rendered (no orphans vs the wireframes).
2. Fidelity + scope — hi-fidelity visuals (not wireframe-grey), no invented structure/nav/token, and it's a seed (not over-built, not claimed production-ready).
3. Visual execution + hierarchy — DS visual language realized; primary element visually dominant; hierarchy legible.
4. Polish — the objective subset + the five named heuristics hold; subjective out-of-set taste is not imposed.
5. Token-backed — every value token-backed (Tailwind/CSS-var/shadcn); no raw off-token value; mapping faithful to the DS.
6. Rendered + self-reviewed — every screen renders without breakage and was vision-reviewed against the wireframe + DS through the loop.
7. Content realism — real content, no lorem; edge data exercised.
8. Rendered states — each flow-visited state rendered, skeleton-not-spinner, error = cause + recovery.
9. Responsive — rendered + verified at the target viewports; mobile content-priority + adaptive nav where form-factor matters.
10. Numeric WCAG 2.2 AA on the render — contrast 4.5/3:1 (SC 1.4.3/1.4.11), focus visible + not-obscured (2.4.7/2.4.11) + appearance ≥2px/3:1 (2.4.13, house rule), target ≥24px (2.5.8), keyboard, reduced-motion; axe-core clean with no rule disabled.
11. DS conformance — no drift/override; a missing token/component surfaced, not inlined.
12. Gaps surfaced — undefined screen/content, missing DS element, a degraded (no-render) run = explicit assumptions.
13. Versioned (on amend) — the doc's own version + changelog reflect the change (Step 8).
14. Capability boundary (n/a when no capability_record) — design-system tokens used throughout; no new components outside the design-system; the rendered scope matches the wireframes scope (interior only for a per-capability record; shell only for system-scope) — no content past a refs entity or an exit_point.

Step 7: Surface gaps explicitly

A wireframe-named screen with undefined content, an element needing a token/component the DS lacks, a contrast/target failure the tokens can't satisfy, or a degraded (text-only, no render) run → an open question or assumption. List it; flag missing tokens/components for the design-system owner. Never paper a gap with an invented token/component or a faked screenshot (§7).

Step 8: Amend mode — edit + re-render the delta, don't redraw (when iterating)

When handed an existing hi-fi + a change, treat it as a scoped, versioned diff, never a regenerate:

1. Scope + ripple — identify the touched screen(s)/state(s), then trace the ripple: every screen consuming a changed DS token through the token→code map (the distinctive wide hi-fi ripple), every screen reusing a changed shared layout/component, every screen added/removed by an upstream wireframe change, every screen naming a changed/deprecated DS component.
2. Edit + re-render the delta — minimal in-place edit; untouched screens' code stays byte-stable + their screenshots are not re-captured (not byte-diffed — headless renders aren't byte-identical run-to-run); re-render + re-vision-review ONLY the delta; a token change is applied once in the map then re-rendered across all consumers in one pass (the one legitimate wide ripple), never hand-edited per screen.
3. Version + changelog — bump the doc's own version: MAJOR (removed/renamed screen or a removed/renamed per-screen state — breaks the build-seed contract), MINOR (added screen/state/component), PATCH (polish/copy/token-value tweak); Keep-a-Changelog entry at the screen/state grain.
4. Deprecate before removing — mark a screen/state deprecated in a MINOR (note the replacement) before removing in a MAJOR; on removal confirm no link/index/sibling reference remains and prune.

Rules

Hard rules (never violate):

• One rendered screen per wireframe-named screen/state. Coverage is keyed to the wireframes; an un-rendered named screen is incomplete.
• Consume the design-system; never invent. Token-backed values only — no raw hex/px; a missing token/component is a flagged gap, not an inline one-off.
• Render the given structure. Lay out the wireframe's screens; never re-decide structure or navigation.
• Run the loop, bounded. Generate→render→screenshot→vision-review→refine before handing off; cap iterations + surface the blocker rather than spin (§9).
• It's a seed, not production. No tests/real backend required; never claim production-readiness; the build phase hardens it.
• Real content, never lorem. Real + edge data; lorem is a defect at hi-fi.
• Numeric a11y on the render, never disable a rule. Fix contrast/focus/target at the token layer; a suppressed axe rule is a finding.
• Vision-capable runtime, or say so. A text-only run is degraded; surface it, don't fake a screenshot.
• Compose, don't duplicate. Take the section structure from the template tool.
• Surface gaps, don't invent. Undefined screen/content, missing token/component, degraded run = explicit assumption.
• Amend, don't regenerate. Edit + re-render the delta + version + changelog; never re-render untouched screens.
• Capability boundary (when capability_record provided). Scope the document output to that record's boundary. Producing content outside the boundary (covering a refs entity as if it were owned; designing flows past an exit_point) is a scope violation — equivalent to inventing content that isn't in the spec.

Preferences (override-able):

• Default standalone HTML+Tailwind; switch to React+shadcn when the build stack is React.
• "Comprehensive" sets ambition but stay proportional — a thin static screen has fewer states + no responsive/interactivity work.
• Prefer a skeleton-that-mirrors-the-layout over a bare spinner where the layout is known.
• The five aesthetic heuristics are judged but named + articulable — don't impose a personal hue/layout preference.

Gotchas

• Blind codegen. Producing code without rendering + seeing it is the #1 failure — the loop (Step 5) is the value; a text-only run is degraded, say so.
• Off-token raw values. A #3b82f6 / 13px that bypasses the token map is drift — bind everything to the DS tokens (Step 3).
• Lorem ipsum. Placeholder text hides overflow/tone/a11y problems and signals an unfinished design — render real content.
• Treating the seed as production. Adding tests/real backends (over-build) or claiming prod-ready (over-claim) — it's a seed; the build hardens it.
• Disabling an axe rule to go green. Excluding color-contrast hides a real defect — fix the token, document a genuine exception.
• Inventing a missing component. A component the DS lacks is a gap for the DS owner, not an inline improvisation.
• Re-deciding structure. Adding a screen/region/nav the wireframe didn't name oversteps into the wireframe's lane.
• Token-change drift on amend. Hand-editing a token's value per screen instead of once in the map is the most common amend regression — change it once, re-render all consumers.

Anti-patterns

• "The code looks right, ship it." UI isn't done until you render + look — run the loop.
• "I'll use a nice blue here." Off-token — bind to the DS token or flag the gap.
• "Lorem ipsum is fine for now." It hides the real layout — use real + edge content.
• "I'll wire the backend + add tests." Over-building a seed — keep it thin; the build hardens it.
• "axe flags contrast; I'll exclude that rule." Fix the token; never disable a real rule.
• "The wireframe is vague, I'll add a screen." Surface it as an assumption instead.
• "It's a small token tweak, I'll regenerate everything." Amend: change the token once + re-render only its consumers + version + changelog.

Output

A comprehensive, rendered high-fidelity UI design as code that meets the Step 6 bar (full coverage vs the wireframes; hi-fidelity + seed scope; visual execution + legible hierarchy; objective + named-heuristic polish; token-backed/no-drift; rendered + vision-reviewed; real content; rendered states; responsive; numeric WCAG 2.2 AA on the render; DS conformance; gaps surfaced; versioned on amend). Each screen is runnable code + the rendered screenshots per viewport × state. The abstract consumer is the human visual gate + the build phase (which grows the seed into production) and a reviewer (which asserts the same bar). The doc's structure comes from the template tool; this skill supplies the content quality.

Worked example (the loop in one screen). Hand-off: a wireframe "Projects — list" (states: populated, empty, loading, error) + a design-system with --primary, --muted, a Card, a Button, a Table. Method: map the tokens into a Tailwind/CSS-var theme (Step 3); build the populated screen with a real Table of representative projects (real names/dates, one long-name row for overflow), the empty state ("No projects yet" + a "New project" Button on --primary), a skeleton mirroring the table for loading, and an error card ("Couldn't load projects — Retry"); render at 360/768/1024/1280, screenshot each state; vision-review → catch that the empty-state CTA is --muted on white at 2.9:1 → fix by binding it to --primary (a token, not a one-off) → re-render → now 4.8:1; axe clean. Result: four rendered states, token-backed, WCAG-passing — a seed the build can grow. (Negative contrast caught by the loop, fixed at the token layer — not by recoloring one button off-token.)

Related

• A hi-fi template tool (a content/template gateway) — supplies the comprehensive per-screen section structure this skill fills.
• agent-browser (+ playwright-best-practices) — renders + screenshots the code for the loop; this skill drives it, doesn't re-teach it.
• shadcn + ui-ux-pro-max / ckm-* — the React component theming + visual-execution guidance; composed-with, not re-taught.
• A deep-research capability — grounds visual + interaction decisions.
• reviewing-hi-fi — asserts the same bar (re-renders + vision-reviews); author + reviewer share one bar so they don't drift.
• The upstream wireframes (a depends_on document) — the screen structure to realize.
• The design-system (a depends_on document) — the tokens + component contracts to consume; it owns the per-component a11y contract, this skill verifies the rendered result conforms.

Progressive disclosure

• references/render-loop-and-tokens.md — the render→vision-review→refine loop procedure + the screenshot-capture protocol (viewport × state) + the DS/DTCG→Tailwind/CSS-var/shadcn token-mapping recipe + the code-stack choice. Load when setting up the loop or the token map.
• references/visual-a11y-and-amend.md — the five aesthetic heuristics ("what good looks like" each) + visual-execution depth + the numeric WCAG 2.2 SCs + axe setup + the amend token-ripple/versioning procedure. Load for a polish/a11y pass or an amend.
• references/sources.md — research provenance for the method + the bar (load only to audit where the guidance came from).

Body budget

• description ≤ 1,024 chars (agentskills.io cap).
• Body ~500 lines / 5,000 tokens (soft target — quality takes precedence; flag if consistently over 700 lines / 7,000 tokens).
• Heavy content lives in references/, loaded on demand.