Design Spec Pipeline (Per-Screen)

Three phases — style-foundation (once) → per-screen spec (loop) → bundle into single upload file. Consumes plans/{slug}/prd/*.md, emits plans/{slug}/design/ with per-screen specs plus one consolidated design-bundle.md ready to upload to any Design AI Tool.

When to Use

• After running kta-prd-pipeline — per-feature PRDs with screen lists exist
• "Generate design specs for my AI design tool"
• "Write Stitch / Figma AI / Claude design prompts for these screens"
• "I need design briefs to feed into a text-to-UI tool"
• Re-running per-screen to refresh after PRD changes

CRITICAL CONSTRAINTS

• SKILL EMITS DESIGN SPECS ONLY — NEVER GENERATE COMPOSE CODE, KOTLIN, OR @PREVIEW STUBS
• COMPONENT-STATES LIST IS MANDATORY — SKILL REFUSES INCOMPLETE SPEC
• REAL COPY ONLY — NO LOREM IPSUM

These rules are absolute. Every reference file reinforces them. If the LLM is tempted to produce Kotlin / @Composable / @Preview, STOP and re-read this section.

How the pipeline works

The pipeline pre-fills sensible defaults (component-state defaults, copy options) and walks the student through each screen step by step. There are no modes, no levels, no toggles — pre-fill, the student picks from options, move forward.

Pre-flight Validation (GATE — runs FIRST, before Phase 1)

Before any phase runs, validate the input. If validation fails, STOP and return — do NOT proceed to Phase 1.

Run these checks in order, in plans/{slug}/:

1. PRD directory exists — plans/{slug}/prd/ is a directory.
2. At least one feature PRD file — at least one prd/{NN}-{feature-slug}.md exists (NN = 2-digit prefix).
3. prd/INDEX.md exists — produced by kta-prd-pipeline Phase 3.
4. Every feature PRD has a ## Screens section with ≥1 screen named in PascalCase.
5. No empty PRD files — each feature PRD is >200 chars (sanity check, not a TBD-only stub).

If any check fails, STOP and emit (no further phases run):

[kta-design-spec] Cannot proceed — input not qualified.

Failed checks:
- {check that failed} — {what was found vs expected}
- ...

Required input shape:
  plans/{slug}/prd/
    INDEX.md
    01-{feature}.md   (with `## Screens` section, screens in PascalCase)
    02-{feature}.md
    ...

Fix path:
  → Run `kta-prd-pipeline` first to generate the PRD set.
  → Or check the slug — passed dir must point to a plans/{slug}/ root.

This skill emits no design specs until input is qualified.

Then return. Do NOT prompt the user to "fix and continue inline" — the upstream skill (kta-prd-pipeline) is the right place to repair PRD shape. The student can re-invoke kta-design-spec after.

Workflow

1. Phase 1 — Style Foundation (once per app) — read references/phase-1-style-foundation.md. Capture mood/voice, color tokens (or "use Material 3 default"), typography, spacing/radii/elevation, references (3 inspirations + 1 anti), motion intent, a11y floor. Save plans/{slug}/design/style-foundation.md. Phase 2 GATES on this file existing.
2. Phase 2 — Per-Screen Spec — read references/phase-2-screen-spec.md. Loop screens across all PRDs. Emit one spec per screen at plans/{slug}/design/{NN-feature-slug}/{NN-screen-slug}.md. Component-states list MANDATORY. Validate before emit.
3. Phase 3 — Bundle (MANDATORY) — after all per-screen specs are written, run scripts/bundle_design_specs.py to concatenate style-foundation.md + INDEX.md + every per-screen spec into a single plans/{slug}/design/design-bundle.md. This single file is the canonical output the student uploads to Stitch / Claude / Figma AI. See "Bundle Step" section below.

There is no per-tool reformatting phase. Modern Design AI tools accept generic markdown; the bundle is tool-agnostic by design. If a specific tool ever chokes on bundle size, paste a single feature folder's screens instead — the per-screen files are still on disk for that.

Bundle Step (Phase 3)

Run after Phase 2 completes — non-optional unless the student explicitly opts out.

Command (use the project venv Python; fallback to python3 if venv missing):

~/.claude/skills/.venv/bin/python3 .claude/skills/kta-design-spec/scripts/bundle_design_specs.py plans/{slug}/design/

The script produces plans/{slug}/design/design-bundle.md containing, in order:

1. Bundle header (title + generation timestamp + upload instructions)
2. style-foundation.md (full content)
3. INDEX.md (full content if present)
4. Each feature directory in NN order, each per-screen spec inside in NN order, with separator comments between sections

After the script runs, surface the absolute output path back to the student in one line, then call AskUserQuestion:

header: "Next Step"
question: "Bundle written to {abs path}. What next?"
multiSelect: false
options:
  - label: "Open in tool", description: "Print upload instructions for Stitch/Claude/Figma AI"
  - label: "Exit",         description: "Stop here — I'll upload the bundle myself"

If the script fails (missing style-foundation.md or no feature dirs), surface stderr and DO NOT proceed to the handoff prompt. Re-run Phase 2 first.

Inputs

• plans/{slug}/prd/*.md (required) — feature PRDs with screens
• plans/{slug}/idea-brief.md (recommended) — for value-prop context

Outputs

• plans/{slug}/design/style-foundation.md — once per app
• plans/{slug}/design/INDEX.md — feature → screens → spec file map
• plans/{slug}/design/{NN-feature-slug}/{NN-screen-slug}.md — one per screen
• plans/{slug}/design/design-bundle.md — single concatenated file ready to upload to Stitch / Claude / Figma AI (produced by scripts/bundle_design_specs.py in Phase 3)

Integration

• Previous stage — kta-prd-pipeline (provides per-feature PRDs)
• Next stage — paste spec into AI design tool (Stitch / Figma AI / Claude artifacts) to generate mockups. NOT into a code generator.

Constraints

• Pre-flight gate — input MUST pass all pre-flight checks (PRD dir, ≥1 feature PRD, INDEX.md, screens section in each PRD, non-stub files). If not qualified, skill returns immediately with a fix path; no design specs emitted.
• Design spec only — no Compose / Kotlin / @Preview / DSL snippets. Reinforced in every reference file.
• Component-states list mandatory — every component must list 6 states (default / pressed / disabled / loading / empty / error). Skill refuses incomplete spec.
• Real copy — every visible string spelled out; no lorem ipsum, no placeholders like {title}.
• Style-foundation gate — Phase 2 refuses to run if style-foundation.md missing.
• Tool-agnostic — Phase 2 output is pure markdown. Phase 3 bundle is the canonical upload artifact for any AI design tool. No per-tool reformatting phase.
• No project-stack-specific tokens leaking in — generic Material 3 defaults are fine; "DRE-KT" / "Compose" mentions forbidden.
• Skill emits markdown only.
• Default working directory — plans/{slug}/.