sketch-implement-design（草图工具设计）

Sketch Implement Design

Overview

Implement UI from Sketch with high visual fidelity by querying live document data through Sketch MCP. Use run_code to inspect and export data, and get_selection_as_image as the visual source of truth.

Prerequisites

• Keep Sketch open with the target document available.
• Enable Sketch MCP in Sketch settings: General -> Allow AI tools to interact with open documents.
• Configure Codex MCP to point to the local Sketch server address shown in Sketch settings.

- Default address is usually http://localhost:31126/mcp.

• Select the layer or frame to implement in Sketch before calling get_selection_as_image.

Required Workflow

Follow these steps in order. Do not skip steps.

Step 0: Confirm Sketch MCP connectivity

If Sketch MCP calls fail:

1. Check that Sketch is running and MCP is enabled in settings.
2. Check the MCP server URL in your MCP client config.
3. Retry after confirming the server address in Sketch settings.

Step 1: Identify the implementation target

Use these approaches in priority order:

• User gives a Sketch share link like https://sketch.com/s/<document-share-uuid>/f/<canvas-frame-uuid>: extract the /f/<canvas-frame-uuid> value, find that layer, and select it. Prefer this target over current selection or name-based matching.
• User gives layer identifiers or names: use run_code to locate matching layers and ask for confirmation.
• User selects the exact frame/component in Sketch.

Use run_code with the find function and this pattern when a share link is provided:

const sketch = require('sketch')
const frameId = '4A2E31FF-56BD-4C29-92D2-829548D19C1D'
const frame = find(`#${frameId}`)
if (frame) sketch.getSelectedDocument().selectedLayers = [frame]

When multiple matches exist, stop and ask which layer to implement.

Step 2: Fetch structured design context via run_code

Run run_code to collect design data for the selected root layer (and relevant descendants):

• Hierarchy: IDs, names, types, visibility, lock state
• Layout: frame, resizing behavior, pins, stack layout, clipping
• Typography: font family, size, weight, line height, alignment, decoration
• Styling: fills, borders, shadows, blur, corner radii, opacity, tint
• Reusable styling: shared text/layer style names and source library names
• Variables/swatches: color variable names and optional source library names
• Symbols: nested symbols and override-capable fields
• Export settings and image sources

Always start scripts with:

const sketch = require('sketch')

Pass a short title argument to run_code and keep scripts minified unless the user asks otherwise.

If output is too large:

1. Fetch a shallow tree map first (IDs, names, types only).
2. Fetch detailed context per critical child subtree.
3. Continue only after all required nodes are covered.

Step 3: Capture visual reference via get_selection_as_image

Call get_selection_as_image on the same selected target. Use this image as the primary visual reference for implementation and validation.

If selection is empty, stop and ask the user to select the intended layer/frame in Sketch.

Step 4: Export required assets via run_code

Export icons, bitmaps, and other assets from Sketch instead of inventing placeholders.

Rules:

• Use assets from Sketch output; do not add new icon packs unless explicitly requested.
• Use absolute output paths.
• For sketch.export, always pass both output and filename (with extension).

Example:

const sketch=require('sketch');const d=sketch.getSelectedDocument()
const l=d.selectedLayers.layers[0]
sketch.export(l,{
  output:'/tmp/sketch-assets',
  filename:`${l.id}.png`,
  formats:['png']
});console.log('ok')

Step 5: Translate Sketch output to project conventions

Treat Sketch data as design intent, then map to project standards:

• Reuse existing UI components before creating new ones.
• Map Sketch values to project tokens (color, spacing, type scale, radius).
• Match established architecture and state/data patterns.
• Keep generated code maintainable and idiomatic for the target stack.

Step 6: Implement for 1:1 visual parity

Aim for visual parity with the Sketch screenshot and context:

• Match spacing, alignment, sizing, and hierarchy.
• Match typography and color usage.
• Preserve intended behavior for responsive/constraint-based layouts.
• Use tokenized values when possible; avoid unnecessary hardcoded values.

Step 7: Validate against Sketch before completion

Validate implemented UI against the screenshot and design context checklist:

• Layout and spacing
• Typography
• Colors and effects
• States and interactions
• Asset rendering
• Responsiveness/constraints
• Accessibility basics

If mismatch remains, re-query with run_code for the affected subtree and iterate.

Implementation Rules

• Prefer incremental updates over broad rewrites.
• Keep component boundaries consistent with the design hierarchy.
• Document intentional deviations (technical or accessibility constraints).
• Do not assume stale design data; re-query Sketch when unsure.

Common Issues

Empty selection from get_selection_as_image

Cause: Nothing selected in Sketch. Fix: Ask the user to select the target layer/frame, then retry.

No document from MCP tools

Cause: No open Sketch document or Sketch not running. Fix: Open the document in Sketch and retry.

run_code output is too large or truncated

Cause: Selection tree is too large. Fix: Query shallow tree first, then fetch detailed context by subtree.

Missing or incorrect exported assets

Cause: Invalid export options or relative output path. Fix: Use absolute paths and pass both output and filename.

Implementation does not match visual design

Cause: Incomplete context capture or token mismatch. Fix: Re-capture screenshot and inspect exact node values via run_code.

References

• Sketch MCP server docs
• Sketch API reference