Figma Pixel Implement

Figma Pixel Implement

Implement UI from a Figma node into the target codebase with export-faithful assets, a screen × state inventory, and a durable living spec in the target repo that figma-pixel-verify can measure against.

Completion boundary: this skill ends at “implemented + measurable contract” (code + file path to Inventory, Spec, and Assets sections). A session-only table is not complete. Missing Assets rows for implemented visible graphics means implement is incomplete. When the project has a component workbench, story coverage of the in-scope inventory (states × context) is also part of implement complete (Workflow 9). Pixel alignment is owned by figma-pixel-verify, which this run continues into. The user-facing result is that skill’s overall PASS. Iterations stay inside the run.

This skill is platform-agnostic (Claude Code / Cursor / OpenCode / others). It requires a Figma design-context capability, not a particular Agent’s skill id.

Visual source of truth: the named Figma node. Walkthrough or QA notes that conflict with the current node are recorded as notes/residuals; they do not replace expected.

Prerequisites

1. Figma design-context available (hard gate) — Confirm the Agent can fetch structured design payload for a node (design context, metadata, screenshot, asset export as the environment provides). If that channel is missing or auth fails, stop: tell the user how to enable Figma MCP / plugin / equivalent for their Agent. Do not invent a single hard-coded install path as the only option. Do not invent UI from the URL alone.
2. Optional: platform Figma→code guidance — If this Agent already ships an official Figma design→code skill or guide (names vary by product; one Cursor plugin example is figma-design-to-code), load it when present and follow its URL parsing / Code Connect / adaptation rules. Missing that named skill is not a blocker—continue with Figma MCP (or equivalent) + this skill’s fidelity rules. Do not require a Cursor-only skill id on Claude Code, OpenCode, or other Agents.
3. Figma URL or node — Prefer a figma.com design/make URL or explicit fileKey + nodeId. If missing, ask once with clarifying-question discipline.

Workflow (ordered)

1. Parse target — Extract fileKey / nodeId (convert node-id=1-2 → 1:2). Note frame name for the spec table.
2. Fetch design context — Call the environment’s design-context tool after step Prerequisites. Treat output as a reference to adapt, not paste-ready final code.
3. Large frames — If the payload is truncated or the frame is clearly multi-section: fetch metadata/structure first, then fetch and implement every visible child section. As each visible section is implemented, append its inventory and spec rows to the living artifact. Unfetched visible sections mean implement is incomplete. Prefer variables/tokens over guessed literals. On design-channel rate limits, pause and report the unfinished sections.
4. Screen × state inventory — Before the spec table, list screens/sections × visual states from the design (component variant names, prototype reactions that map to appearance, and readable annotations / comments / adaptation notes). Documented states such as default / hover / pressed / disabled / selected become inventory entries. Readable notes become inventory rows (blocked, visual state, or out-of-scope: behavior as they fit). Prototype or product behaviors that are not computed-style rows (window drag, auto-follow scroll, mouse-wheel paging) stay on the list as out-of-scope: behavior. If the design has no Hover (or equivalent) variant, record no-variant-in-design — do not invent hover colors or sizes from a note. When a note conflicts with the named node’s structured value, keep expected on the node and record the conflict. When the component’s host container can vary (resizable width, narrow squeeze), add a context axis to the inventory: the product width floor, preset checkpoints, and squeeze pressure become inventory rows or an optional context column.
5. Detect theme structure — Establish how the design expresses themes: variable collections with multiple modes (e.g. light/dark), or separate per-theme frames/variants. Theme is in scope when the file has multiple modes or the project already switches theme (data-theme, token files, prefers-color-scheme). Record the theme inventory (mode names or per-theme frame nodeIds). When neither the file nor the project has a second theme, stay on the requested frame/variant.
6. Build the living spec — Resolve a durable path in the target repo: prefer an existing Figma doc directory if the project already has one; otherwise docs/figma/<safe-frame-name>.md. Create Inventory, Spec, and a required Assets subsection there (leave Verify empty for figma-pixel-verify). After the inventory, write measurable rows — mechanically expanded from the fetched node tree, never sampled. Every fetched visible node yields rows for its position (x/y relative to its parent or section) and size; every text node additionally yields its full typography set (font-size, weight, line-height, color, alignment) and its line count when multi-line; every container additionally yields padding, gap, radius, shadow, and each child's cumulative absolute offset from the frame origin; sibling spacing is recorded with basis naming the relationship. A fetched node is either covered by spec rows or listed as out-of-scope with a reason — silent omission makes the spec incomplete. Minimum columns: node / role, state, property, expected (Figma exact), unit, basis (which parent or sibling the number is relative to), token / class in repo, source (variable, MCP field, metadata, or Code Connect component), primitive (project-component | custom-layout | platform-control). Fill ownership: if the node has no readable fill, Spec expected is transparent / none — a screenshot of surrounding color does not own that fill. Text color: use the text node’s own readable Figma variable as token / source; if none, record raw-only. Screenshots are visual reference only—not the sole source of numbers. When theme is in scope, add a mode column (or one row per mode) for theme-varying properties: resolve per-mode values from the design itself (variable valuesByMode when the channel exposes it, otherwise re-fetch with the file/frame mode switched or via the per-theme frame); shared values stay single-row. A session note is a draft only—implement is incomplete until the file exists (including Assets). Reference exports: export a raster image of every in-scope frame from the design source, store it in the target repo, and record node id, local path, and export scale in the living artifact. figma-pixel-verify uses these as its differential-sweep baseline; a missing reference export for an in-scope frame means implement is incomplete. Template: reference.md.
7. Map to project — Reuse existing components, tokens, and layout primitives. Prefer design-system variables over hard-coded values when the project already has them; flag unbound one-offs; when hard-coding is required for fidelity, record the expected value in the spec table. Layout-measurable control: when the node publishes an explicit width and height, the control that remains after layout inside the shipping parent (the container the screen uses) must have that box. A probe that lays out the control with no parent does not count. Prefer a project component whose laid-out size matches. When every stock platform control lays out to a different size, build a custom control and assign track, thumb, and label frames in that layout pass (primitive: custom-layout). platform-control is valid only after a layout probe of a stock instance matches the spec box. A flattened payload (outer box only) still binds that outer box to a matching project component or custom-layout. Inner sizes come from a structured child fetch or an existing spec row that cites a design node.
8. Assets (hard rules) — See reference.md whitelist/blacklist. Export real image/SVG bytes into the project (or an approved dynamic source). Cut each visible icon/logo as one file from the outer instance (not stacked leaf paths). Reconstruct that instance’s width, height, and offsets — not a generic 16/20/24 shell unless the instance uses it. Restore overlap and floating layers from the readable instance hierarchy. Keep SVG aspect ratio unless the design requires stretch. Use <img> / framework image / export-preserving SVG. Do not use hand-authored path placeholders, CSS mask / mask-image + fill/currentColor recolor, or invent mask theming when only a light export exists (prefer a second export or record pending). When theme is in scope, export design-provided assets per theme (from each mode or per-theme frame)—one theme's export plus an invented recolor stays forbidden. Record every in-scope graphic in the Assets subsection (node, local path, format).
9. Stories & component tests — When the target project has Storybook or a similar component workbench (existing story setup, workbench config, or scripts): write stories covering every in-scope Inventory row (state × context), each exposing its state through a declarative entry the Agent can drive later (test id, knob, or auto-action), following the project’s existing story conventions. Add a Stories subsection to the living artifact mapping story → covered inventory rows → entry (template: reference.md). Write component tests locking the key Spec values (geometry, color, token mapping) when the project has a test setup; when it does not, record the suggestion instead of scaffolding (advisory, not blocking). When no component workbench exists, record the adoption suggestion once (isolated, observable, agent-testable surface) — measurement falls back to the channel ladder in figma-pixel-verify.
10. Theme mapping — Stay design-faithful to the requested frame/variant colors. When theme is in scope, map theme-varying variables onto the project's theming mechanism (CSS custom properties / design tokens / data-theme or prefers-color-scheme switching) instead of duplicating hard-coded per-theme literals.
11. Hand off — Give the living-artifact path (Inventory + Spec + Assets present; + Stories when a workbench exists). In the same run, load figma-pixel-verify and follow it. On a non-pass row, that skill fixes and remeasures before the user sees a result. The user-facing result is overall PASS, or FAIL that names a measurement command that could not run. No path = implement incomplete.

Tool intent (platform-agnostic)

Describe goals; let the Agent pick native tools:

Do not require a named MCP server id or CLI as the only path.

Relationship to other skills

Pitfalls

• Treating a Cursor-only skill name as required on every Agent.
• Treating MCP code as final without project adaptation.
• Claiming “pixel perfect” without figma-pixel-verify.
• Using CSS mask recolor for icons that Figma exported as flat assets.
• Skipping the spec table so verify has nothing measurable.
• Treating a session-only table as implement complete (no durable path).
• Sampling a large frame’s chrome and calling implement complete.
• Inventing hover or dark values the design does not provide.
• When theme is in scope: implementing only the requested theme and hand-rolling the other themes' values or assets instead of resolving them from the design.
• Using the node’s own padding as expected when the design relationship is “distance to sibling X”.
• Claiming pixel PASS for out-of-scope: behavior items.
• Painting a solid fill because a screenshot looks filled when the node has no readable fill.
• Reassembling an icon from leaf-path fragments or a generic 16/24 wrapper.
• Replacing a readable Figma text variable with a self-invented semantic alias.
• Treating implement complete without an Assets subsection.
• Declaring implement complete when a component workbench exists but stories do not cover the in-scope inventory rows.
• Writing stories/tests that assert numbers from chat or walkthrough notes instead of the Spec table’s expected.
• Sampling “key values” into the spec table instead of mechanically expanding every fetched node — a missing row is unmeasurable downstream (unknown unknown).
• Handing off without per-frame reference exports — the differential sweep has no baseline.