Designing

Designing — Design Tool Integration

Route to the right design tool based on the URL the user provides — Paper (preferred), Stitch, Figma, or local assets — to extract layout, content, and visual references for implementation.

When to use

• Starting implementation of a visual design
• Need to capture design reference before building
• Comparing implementation against design
• Populating plan assets directory with screenshots

Inputs required

• Design source: Stitch project ID, Figma file URL, or local asset path
• Optional: specific section/screen name to extract

Procedure

0) Determine source from user input (URL-based routing)

Routing is driven by what the user provided, not by which MCPs happen to be configured.

1. URL match — inspect the user's input for a design URL:
   • paper.design/* or *.paper.design/* → paper branch (use mcp__paper__*)
   • figma.com/* → figma branch (use mcp__figma__* / mcp__claude_ai_Figma__*)
   • stitch.withgoogle.com/* (or other known stitch hosts) → stitch branch (use mcp__stitch__*)
2. Path match — inspect user input for a .pen reference:
   • Input ends in .pen OR input starts with design/ → pencil branch
   • No input provided AND design/ exists in the project root with .pen files: list the available .pen page files (exclude *.lib.pen) and ask the user which page to extract before proceeding.
3. Local fallback — if no URL but docs/plans/<active-plan>/assets/section-*.png exists → offline branch
4. Ask — if neither URL nor local assets are present, ask the user for one

MCP gate: once the branch is known, ToolSearch the corresponding mcp__<tool>__* namespace. If the MCP is NOT configured, stop with this message:

⛔ You sent a {paper|figma|stitch} link but the `{tool}` MCP is not configured.

Configure it and re-run, or send a link from another source.

Do NOT silently fall back to a different MCP.

Pencil MCP gate: For the pencil branch, ToolSearch mcp__pencil__open_document. If NOT configured:

⛔ .pen file detected but the Pencil MCP is not configured.

Install: claude mcp add pencil -- npx -y @anthropic/pencil-mcp
Restart the session after installing.

1) Extract design data (per section, never full design at once)

Paper workflow (preferred when source is paper.design):

1. mcp__paper__get_basic_info — get document metadata
2. mcp__paper__get_tree_summary — locate the target section node
3. mcp__paper__get_node_info on the section — capture structure, text, hierarchy
4. mcp__paper__get_screenshot — save as assets/section-{name}.png
5. mcp__paper__get_computed_styles — save as assets/section-{name}.styles.json (typography, colors, spacing — exact values; consumed by verifying for the style spot-check)
6. mcp__paper__get_jsx — save as assets/section-{name}.reference.jsx with this header comment as the FIRST lines of the file:

   // REFERÊNCIA ESTRUTURAL APENAS — NÃO COPIAR.
   // Sage usa Blade, não React. Use isso só para entender
   // hierarquia de componentes e nesting.
   

7. Produce the structured output (see step 2) — same schema as the other branches.

Stitch workflow:

1. mcp__stitch__list_projects — find the project
2. mcp__stitch__list_screens — enumerate all screens
3. mcp__stitch__get_screen — extract one section at a time
4. For each section, capture: headline, body text, components, colors, layout structure

Figma workflow:

1. mcp__figma__get_file — load the file structure
2. Navigate frames to find sections
3. Extract text layers, colors, component structure per section

Pencil workflow:

1. open_document(filePath) — open the .pen file the user indicated
2. get_editor_state() — confirm top-level nodes and document is active
3. batch_get with no nodeIds, readDepth: 1 — map all available sections
4. For each section to extract: delegate to pencil-extractor in SURGICAL mode, passing filePath, sectionId, and planPath
5. After all sections: optionally invoke pencil-extractor in COMPONENT_MAP mode to produce design/component-map.md

Offline workflow:

1. Read images from docs/plans/<plan>/assets/section-*.png
2. Claude reads images natively — describe layout, content, colors
3. If no assets exist, ask user to provide screenshots

2) Structure the output

For each section extracted, output:

### Section: {name}

**Layout:** {grid structure, column arrangement, alignment}
**Headline:** "{exact text}"
**Body:** "{exact text}"
**Components:** {cards, buttons, icons, images — with details}
**Colors:** {background, text, accent — hex values if visible}
**Typography:** {heading size, body size, weight, font family}
**Spacing:** {padding, margins, gaps — approximate}
**Icons:** {icon names/types, from which set}

3) Save to plan assets (if active plan exists)

If there's an active plan in docs/plans/:

• Save extracted data as structured notes in assets/section-{name}.md
• If screenshots are available, note their paths
• Update plan.md frontmatter with design-tool: paper|stitch|figma|offline

Key Principles

• Granular extraction — always per-section, never full design at once (prevents context overflow)
• Exact content — copy text verbatim, don't paraphrase headlines or body text
• Persist to disk — save everything to plan assets/ so it survives context compression
• Re-read before implementing — always re-read assets from disk, never rely on context memory