-
Notifications
You must be signed in to change notification settings - Fork 7
All 22 Commands
Every command ux-skill ships, organized by workflow stage. Each command can be called directly from Claude Code, Cursor, Windsurf, or any of the 17 supported IDEs.
DISCOVERY → RECOMMEND → GENERATE → QUALITY → WORKFLOW
Each command sits at one of these stages, and the engine chains them via .ux/last-*.json state files.
The forcing function. Runs a 10-field discovery protocol that refuses to start generation until the brief is locked. Replaces "make a landing page" with a structured brief that has industry, audience, tone, must-haves, forbidden, stack, region, success metric, reference brands, and primary goal.
ux discover
# Saves: .ux/last-discovery.jsonWhy it exists: most LLM design output is bad because the model jumps to a default aesthetic instead of reading the room. Discovery makes the LLM ASK before it generates.
The flagship. Runs the 5-parallel-search engine across 11 manifests:
Industry → Style → Palette → Type → Motion + Components + Brand Exemplars + Guardrails
Returns a complete recommended design system in one JSON document.
ux recommend --brief-file .ux/last-discovery.json
# Saves: .ux/last-recommendation.jsonReads:
-
data/industries.json(184 entries) -
data/styles.json(84 entries) -
data/palettes.json(176 entries) -
data/type-pairs.json(70 entries) -
data/motion-presets.json(57 entries) -
data/components.json(148 entries) -
data/brands/*.json(110 entries) -
data/anti-patterns.json(100 entries, always-on guardrails)
The reasoning is deterministic — same input always produces same output. Industry → recommended_styles + avoid_styles biases the style lane. Each lane scores entries against the brief's tone/audience. The top picks are merged into one recommendation with an explainable rationale.
See: Discovery Protocol, Recommender Engine.
Generate a complete surface (landing, dashboard, app screen) from a recommendation + a section brief.
Generate a single component (button, card, table, form) grounded in the recommended system.
Build the entire design system from scratch — tokens, foundations, components, theme layers — when no system exists.
Generate a data-dashboard system (cards, tables, charts, filters, empty/loading states).
Discovery → recommendation → component in one chained pass. Fast path for a single deliverable.
Run the 100-rule deterministic regex linter over file paths. No LLM in the audit lane. Exits non-zero on Critical/High findings — CI-friendly.
ux lint ./src --threshold highSee: 100 Anti-Pattern Catalog.
LLM-driven 6-lens review (taste / a11y / motion / content / brand fit / performance). Slower than /ux-lint, deeper than a regex.
See: The 6 Audit Lenses.
Accessibility-focused audit. WCAG 2.2 AA checklist + heuristic review.
Design critique. Treats the artifact as a finished piece to be reviewed in a senior-designer voice. Flags taste calls + asks pointed questions.
Apply /ux-lint findings inline. Reads .ux/last-lint.json, rewrites the source file with fixes.
LLM-driven taste pass. Refines microcopy, spacing, motion easing, color balance. Less aggressive than /ux-fix — closer to "polish the demo before show & tell".
Microcopy generator. Writes button labels, empty states, error messages, success states, onboarding tooltips, terms-and-conditions language. Uses the recommended brand voice from .ux/last-recommendation.json.
Motion preset generator. Pulls from the 57 motion presets, picks the right easing curve + duration for the use case (entry / hover / transition / scroll / state / exit / attention / press).
Distill research notes (interviews, analytics, competitive sites, A/B results, support tickets) into actionable design recommendations.
Facilitate a remote design workshop. Generates the prompt sequence, the Miro board structure, the synthesis frame.
Write a case study from product data. Pulls metrics + screenshots + voice + brand spec to assemble a Polaris-style write-up.
Suggest the next milestone. Reads the recommendation + the current artifact state, returns "ship X next, then Y, then Z."
Load an LLM with "expert mode" — deeper reasoning, slower, more taste calls. Use when stuck on a hard design decision.
Engine telemetry. Returns version + per-manifest counts.
ux stats
# {
# "version": "2.0.0-alpha",
# "counts": {
# "styles": 84,
# "palettes": 176,
# "type-pairs": 70,
# "components": 148,
# "industries": 184,
# ...
# }
# }Print the MCP server config (claude_desktop_config.json snippet, Cursor settings, Windsurf MCP config) to register ux-mcp in your MCP host.
See: MCP Server.
Every command writes its output to .ux/last-<cmd>.json. The next command reads from it. The full chain:
ux discover # .ux/last-discovery.json
ux recommend --brief-file .ux/last-discovery.json
# .ux/last-recommendation.json
ux design hero # generates Hero.tsx from recommendation
ux lint ./src --threshold high # .ux/last-lint.json (exits non-zero on issues)
ux fix # applies fixes inline
ux polish # final taste pass
ux persist save --project-root . # writes .ux/design-system/MASTER.mdAfter ux persist save, anyone cloning the project can ux persist load and the same recommendation re-renders. The state is reproducible.
DISCOVERY: ux-discover
RECOMMEND: ux-recommend
GENERATE: ux-design ux-component ux-system ux-dashboard ux-frame
QUALITY: ux-lint ux-audit ux-a11y ux-critique ux-fix ux-polish
WORKFLOW: ux-copy ux-motion ux-research ux-workshop ux-case-study
ux-next ux-expert ux-stats
META: ux-mcp
22 commands total. Every one is documented at uxskill.laithjunaidy.com/commands.html.