---

name: plan-uiux-unification description: > Plan-only UI/UX and design-system audit that emits a unification burndown; no code until a phase is approved. Use when "audit UI without fixing", "UI burndown", or "IA audit before redesign". Apply now → housekeep-design. license: MIT effort: high

UI/UX & Design-System Audit + Unification Plan

Degree of freedom: MIXED — IA and unification plan are judgment; token greps and inventory globs run exactly. Stay plan-only. No UI edits until approved.

Role: Senior frontend/mobile architect + design-systems engineer.

Task: Deep, exhaustive, non-destructive audit. Built by many devs over time → expect inconsistent design system, layout, typography, and repo rules. Find every violation, document it, produce a burndown + enhancement plan. Enhance the existing design system; do not replace it, strip features, fabricate data, or rewrite UI in this pass.

This skill vs neighbors

Skill	Does
plan-uiux-unification (this)	Plan only — full inventory, burndown, roadmap; no fixes
audit-responsive	Linearized desktop / breakpoint IA — audit then fix
audit-uiux-design-system	Visual/token/component audit; may recommend or fix
audit-ux	Heuristics, flows, microcopy quality
enhance-web-ui / enhance-web-ux	Implement visual/UX improvements
design-system	Build or extend the design system itself

Chain: Step 1 of six-skill plan loop — see docs/PLAN-LOOPS.md
Execute: audit-responsive (layout/IA) → enhance-web-ux / enhance-web-ui → test-playwright

How to reason (every plan item)

1. Propose — token, primitive, layout, or voice change against the existing DS
2. Risk — brand-break, a11y fail, or a rewrite that strips a feature
3. Keep-working — surfaces and tokens that already match the canonical system
4. Phase — P0 broken/a11y → structural → polish (do not execute)

Worked example

Propose: replace raw #1a1a1a on /pricing CTA with the documented primary token (bg-fg or equivalent). Risk: 14 hex literals on the highest-traffic page; brand-breaking vs the token SSOT. Keep-working: components/ui/Button already uses semantic tokens. Phase: P0 — brand-breaking hex on a conversion surface. Canonical fix: value-for-value token swap; no copy or layout rewrite.

---

⛔ Preservation Contract

Read references/preservation-contract.md in full before Phase 0.

Acknowledge in output #1 — restate what you will NOT do:

• No feature/route/screen/state/prop/handler removal
• No fabricated data, content, copy, or API fields
• Every codebase claim cites a file path you read
• Preserve I/O boundaries + business logic unless explicitly proposed + approved
• No silent breaking changes to public APIs, props, or routes
• This pass plans; visual/UX changes are proposals only

---

Phase flow (do not skip) [HIGH freedom]

0. Context & IA (before any layout talk)
1. Design-system ground truth
2. Full surface inventory
3. Per-surface deep audit
4. Burndown (prioritized + risk)
5. Unification & enhancement plan
6. Best-practice research (cite current year)

Detailed instructions: references/phase-workbook.md
Templates: references/output-templates.md

---

Phase 0 — Context & IA FIRST [HIGH freedom]

Before pixels: user goal, primary task, key action, IA map (primary/secondary/tertiary), content relationships, progressive disclosure.

Core anti-pattern to flag: flat vertical stacks + left-align + no hierarchy/grouping.

Only then propose layout using real content from the codebase — never invented examples.

---

Phase 1 — Design-system ground truth [HIGH freedom]

Auto-detect (read actual files): [LOW freedom — run exactly]

Glob: **/tailwind.config.*
Glob: **/globals.css **/tokens.* **/theme.*
Glob: **/components/ui/**
Grep: "STYLEGUIDE|CONTRIBUTING" glob "**/*.md"

Critique: enhance (missing semantic tokens, thin scales) vs enforce (violations of existing rules). Multiple systems → name canonical vs legacy + collision points.

Output short design-system spec (current state + proposed enhancements).

---

Phase 2 — Surface inventory [HIGH freedom]

Exhaustive checklist table — every route, modal, drawer, toast, empty/loading/error state, auth/onboarding, responsive + dark per surface. Mark audited only after Phase 3 review.

Delegate an inventory pass to a subagent only when it is sizeable, independent of the others, and its raw output (glob lists, token-grep counts) has no place in the plan; brief it with the exact commands and the columns to return. It edits nothing, like this pass. Long sweeps are expected — context compaction exists, so finish the inventory rather than wrapping up early.

See inventory template in references/output-templates.md.

---

Phase 3 — Per-surface audit [HIGH freedom]

For each surface, log violations by category:

• IA & hierarchy · Tokens · Typography · Layout/spacing
• Content & tone (rewrites preserve meaning)
• AI-generic aesthetic tells
• Components (duplicates → consolidation proposals)
• Interaction/states · Accessibility (WCAG 2.2)
• Conflicting/deprecated systems · Tech debt (removal needs confirmation)

Three buckets: violates documented rule · subjective improvement · needs DS enhancement

Token sweep (quantify for burndown): [LOW freedom — run exactly]

Grep: "#[0-9a-fA-F]{3,8}" glob "*.{tsx,jsx,vue,css}"
Grep: "style=\\{" glob "*.{tsx,jsx}" output_mode "count"
Grep: "className=.*\\b(p|m|gap|text|rounded)-\\[" glob "*.{tsx,jsx}" output_mode "count"

Optional visual verification: read protocol-browser-anti-stall and references/playwright-session-coordination.md in that skill before playwright-cli. Plan-only — screenshots for evidence, not fixes.

---

Phase 4 — Burndown [HIGH freedom — plan only]

Table columns:

Surface | Violation | Category | Severity (P0–P3) | Effort (S/M/L) | Risk (Low/Med/High) | Canonical fix | File path(s)

P0 = broken/inaccessible/brand-breaking · P3 = cosmetic. Quantify where possible.

---

Phase 5 — Unification plan [HIGH freedom — plan only]

• Consolidate DS conflicts; deprecate loser + migration path
• High-traffic layouts driven by Phase 0 IA + real content
• Content/voice guideline (tone, errors, CTAs, jargon glossary)
• Token migration / codemod strategy (value-for-value)
• Primitive consolidation as proposals with dependency checks
• Guardrails: lint rules, no raw hex, visual regression, PR checklist
• Phased roadmap: quick wins → structural → polish (each revertible)

---

Phase 6 — Research [HIGH freedom]

Firecrawl / web for current year best practices — search each named standard as written rather than answering from memory; cite sources:

• Token architecture (semantic/layered, W3C DTCG)
• IA + visual hierarchy · UX writing · anti-AI-slop aesthetics · WCAG 2.2

Note where this app diverges from research.

Self-critique before the burndown [LOW freedom — do not skip]

1. evidenced-not-assumed — every violation cites a path you read; no invented copy
2. plan-only — no UI/token edits, no PRs
3. severity/phase justified — P0 is broken/inaccessible/brand-breaking, not taste
4. right-owner — linearized breakpoint IA → audit-responsive; apply-now token cleanup → housekeep-design; per-page heuristics → audit-ux
5. no-false-safety — [NEEDS REAL CONTENT] stays open; never fabricate examples to fill IA

---

Required output (in order)

1. Preservation-contract acknowledgment
2. Context & IA summary (per area)
3. Design-system spec (current + enhancements)
4. Surface inventory checklist
5. Per-surface violation log
6. Burndown table (with risk column)
7. Unification + enhancement plan + phased roadmap (+ content/voice guide)
8. Guardrails / tooling recommendations
9. Research notes + citations
10. Open questions / [NEEDS REAL CONTENT] list

Deliver as a single markdown doc (or canvas for large repos), then end the turn with a standalone recap in chat: the two or three highest-impact violations and the first phase to approve. The document is the deliverable — write it before the recap. Do not open PRs or edit UI code unless the user explicitly approves execution in a follow-up.

---

Rules

• Plan only — do not fix yet. User approves before execution.
• IA before layout. Hierarchy first, pixels second.
• Enhance, don't replace. Preserve every feature + all real content/data.
• Never fabricate. Cite real paths/lines. Unverifiable → say so.
• Surface unreachable or token source missing → say so, don't guess.
• Before proposing any file change: one line on what must keep working there.

---

MCP integration

MCP	Use
Firecrawl	Phase 6 research, anti-AI-slop patterns, WCAG/token governance
Playwright browser	Optional Phase 3 visual evidence; own named session (-s=…) per protocol-browser-anti-stall
Context7	Component library / Tailwind docs when naming canonical primitives
Sequential Thinking	Not needed — the IA map and collision analysis go in the deliverable