prd-to-ux

prd-to-ux

Take a PRD (or a raw product idea) and produce tool-agnostic, concept-driven screen prompts for web UX generators — Stitch, Figma AI, Pencil, Claude Design, and similar.

Output describes design intent: what a screen is for, what's visible, how it feels. It does NOT emit Tailwind, component trees, or hex codes (unless the user explicitly supplies them).

Pipeline

• Step 1 — PRD Shaping (optional) — when no PRD exists. See references/step-1-prd-shaping.md.
• Gate — PRD Validation — required pass between Step 1 and Step 1.5. See references/prd-validation.md.
• Step 1.5 — UX Philosophy (required) — 3 distinct philosophies, user picks. See references/step-1-5-ux-philosophy.md.
• Step 2 — Feature & State Briefs (required) — screens and states per feature. See references/step-2-feature-states.md.
• Step 3 — Design Direction (optional) — vibe brief, or shadcn-style default. See references/step-3-design-direction.md and references/default-design-direction.md.
• Step 4 — Screen Prompts Export (required) — one prompt block per screen × state. See references/step-4-screen-prompts.md.

The divergent→convergent pattern shared by Step 1 (features) and Step 1.5 (UX expressions) lives in references/divergent-convergent-pattern.md.

Artifacts

Written to the user's current working directory under:

docs/prd-to-ux/<YYYY-MM-DD>-<project-slug>/
├── 01-prd.md
├── 01b-ux-philosophy.md
├── 02-features-and-states.md
├── 03-design-direction.md        (only if user opted in at Step 3)
└── 04-screen-prompts/
    ├── <feature-1>.md
    └── ...

Project slug is lowercase kebab-case derived from the PRD title or elevator pitch. Skill proposes a slug and lets the user override.

Entry

User invokes /prd-to-ux or references the skill, optionally passing a PRD file path or pasted content.

Opening Move

1. Introduce the pipeline stages briefly (Step 1 through Step 4, plus the validation gate).
2. Ask: "Do you have a PRD, or should we build one from an idea?"
3. Check the target directory docs/prd-to-ux/<today>-<proposed-slug>/ for existing artifacts.
   • Directory exists with artifacts: ask whether to resume at the next missing step or start fresh.
   • Directory does not exist: propose the slug, confirm, then create it and proceed.

Flow Control

• One step runs at a time. Between steps, show a terse summary (what was produced, where it lives) and ask to proceed or refine. Refinement edits the artifact in place before moving on. This is the phase gate — do not batch steps.
• Artifacts are written incrementally within each step. A dropped session can be resumed on restart.
• Skill never auto-selects at any convergence point. User chooses the philosophy, the design direction, the feature scope. See the individual step references for specifics.

Termination

The skill stops after Step 4 (Screen Prompts Export). It does NOT invoke Stitch, Figma AI, Pencil, Claude Design, or any other rendering tool. The user takes the prompts elsewhere.

Edge Cases

• Vague feature at Step 2: pause, ask targeted questions, update 01-prd.md, resume. Never fabricate.
• User pastes a PRD mid-flow: write verbatim to 01-prd.md, run validation gate, skip divergent-adversarial at Step 1, proceed to Step 1.5.
• User rejects all 3 philosophies at Step 1.5: re-diverge with different constraints, or take user-supplied axes. Do not force a choice.
• Design-direction vs. functional tension (e.g., "minimal" + dense data table): surface in the relevant screen prompt as a note to the rendering tool, don't silently resolve.
• Platform mismatch (mobile PRD, user plans to paste into a web-focused tool): note in final handoff; suggest specifying platform framing when pasting.

Out of Scope

• Invoking rendering tools on the user's behalf.
• Generating code (React, Tailwind, component files).
• Pixel-level style specs (hex, dp, spring tensions) — unless the user supplied them in Step 3.
• HTML preview artifacts, consultant-posture mockup proposals.
• Iteration on rendered output (belongs to the rendering tool).
• Maintaining a cross-project design system.

Red Flags / Common Rationalizations

Thought	Reality
"Skip the philosophy step, the user knows what they want"	Step 1.5 is required. The 3-philosophy frame is where expression-level coherence is decided.
"Auto-pick the recommended philosophy — user seems decisive"	Never auto-select. Always wait for the user's choice.
"Skip empty/error states to move faster"	No. Every state gets a prompt.
"Add Tailwind classes to make prompts more actionable"	The skill is explicitly tool-agnostic. Describe, don't prescribe.
"Batch multiple steps in one message to save turns"	Phase gates exist to catch drift early. Run one step at a time.