# Audit Design System

Review a Figma node for evidence that the design is not properly integrated with the design system.

This skill is read-only. When the user wants a write action afterward, downstream skills should use `use_figma` through a `figma-use`-style helper when the host environment requires one.

## Output Format Selection

- **Explicit user request wins**:
  - If the user asks for `--json` or JSON, output raw JSON (no markdown fences, no prose).
  - If the user asks for `--markdown`, markdown, or a specific human-readable format, output the human-readable markdown report.
- **Codex Desktop app**: Output raw JSON by default.
- **Codex CLI and other chat-style environments**: Output the human-readable markdown report by default.
- **Machine-consumed review surfaces**: Output raw JSON by default.
- **Ambiguous environment**: If the environment is unclear, output markdown by default.

## Workflow

1. Parse the Figma input.
   Accept a full Figma URL, or a `fileKey` and `nodeId`.
   Normalize node IDs from `72-293` to `72:293` when needed.

2. Pull the minimum required evidence with Figma MCP read tools.
   Call `get_design_context` for the exact node under review.
   Call `get_screenshot` for visual confirmation.
   Call `get_variable_defs` to see which variables are actually bound.
   Call `get_code_connect_map` when relevant.
   Call `get_metadata` when the reviewed node is large, repeated, or board-like and you need to map nested instances before drilling in.
   Call `search_design_system` when you have identified a likely non-systemized primitive and there is a realistic chance of suggesting a concrete replacement from the audited design system.

3. Review for systemization failures, not visual taste.
   Look for places where the design should probably inherit from the design system but is locally constructed instead.
   Base every finding on structure visible in Figma: instances, duplicated frames, raw values, variant drift, or missing token bindings.
   Prefer omissions over weak findings.

4. When the evidence is strong enough, suggest a replacement candidate.
   After identifying a likely custom primitive, use `search_design_system` to find the closest matching component family from the audited design system.
   Include a candidate only when the match is credible from structure and naming, not just screenshot similarity.
   If search results are noisy or ambiguous, omit the candidate instead of guessing.

5. Present findings in the appropriate format based on the environment.
   Use JSON for Codex Desktop and machine-consumed review surfaces, markdown for Claude Code CLI and other chat-style environments (see Output Format Selection above).

6. When the user wants a fix, route to the right downstream skill.
   Prefer [fix-design-system-finding](../fix-design-system-finding/SKILL.md) when one specific offending node should be repaired.
   Prefer [apply-design-system](../apply-design-system/SKILL.md) when the user wants a broader screen-wide pass, multiple sections need coordinated remediation, or the review is being used to define scope before writing.

## What To Flag

- Shared UI primitives recreated as ad-hoc frames instead of component instances.
  Common targets: buttons, icon buttons, cards, alerts, pills, chips, avatars, stat tiles, tab bars, nav bars, FABs, list rows.

- Repeated sibling structures that should clearly collapse into one reusable primitive.
  Example: three nearly identical stat tiles with different content.

- Hard-coded visual values where the rest of the design system uses variables.
  Common targets: fills, strokes, text colors, radius, spacing, typography, shadows.
  Only flag this when the evidence is concrete, such as a raw hex value or bespoke geometry sitting beside tokenized peers.

- Global navigation or other high-leverage patterns built from custom frames instead of system components.
  Flag these aggressively because drift there scales across many screens.

- Variant drift inside a nominal component.
  Example: a local edit button with unusual size, stroke width, or radius that does not match the expected icon-button primitive.

## What Not To Flag

- Purely aesthetic preferences.
- Copywriting or product decisions.
- Layout choices that can reasonably remain screen-specific.
- One-off compositions when the underlying primitives are already componentized and tokenized.
- Claims that require undocumented assumptions about a design library.

## Evidence Standard

Every finding must answer both questions:

1. What concrete Figma evidence shows this is not systemized correctly?
2. Why does that matter for propagation, consistency, theming, or maintenance?

Good evidence includes:

- a node is a plain frame when it should be an instance
- several siblings duplicate the same structure
- raw color or geometry values appear where variables or standard primitives should apply
- a global pattern is custom-built

Weak evidence includes:

- "this looks custom"
- "I

…