Figma Design System Builder

The evidence shows discovery-first execution, explicit user approval checkpoints, strict sequencing, returned node IDs, and a prohibition on unconfirmed destructive cleanup; there is no evidence of credential theft, covert exfiltration, or malware. Points are deducted because the skill performs extensive external writes to Figma, rollback is limited to cleanup and state tracking, and remote MCP data flow, permission boundaries, and recovery responsibilities are not fully disclosed.

The evidence shows a structured phased workflow, state ledger, idempotency checks, validation steps, and an error-recovery reference. Points are deducted because the main document requires return and forbids closePlugin, while the component reference still uses IIFEs, closePlugin, and closePluginWithFailure; the main document says getPluginData is unsupported, while reference code uses setPluginData. The helper scripts and key dependency contents are not supplied, so static review cannot reproduce critical paths and the score remains at or below 10.

The evidence defines the audience, use cases, inputs, workflow, and boundaries against prematurely building components reasonably clearly. Points are deducted because figma-use must also be loaded and core functionality depends on the Figma MCP; Chinese-language interaction is not addressed, and reachability of https://mcp.figma.com/mcp from mainland China is unverified.

The evidence provides phase-based information architecture, reference documents, script listings, naming rules, limitations, examples, and dependency notes. Points are deducted because license metadata is unknown, and no versioning policy, changelog, or clearly assigned maintenance responsibility is shown; maintainers.yml pointing to mcp_server is insufficient evidence of governance or an update path.

The evidence covers tokens, components, variants, documentation, Code Connect, and QA, with checkpoints intended to support a complete design-system build. Points are deducted because no committed test suite, CI execution evidence, or representative verified outputs are supplied, and some examples contain API conflicts; under static calibration, the task design is plausible but direct usability is unverified, so the score remains at or below 7.

The evidence includes a detailed workflow, code snippets, parameter descriptions, and error-code tables, giving it some auditability. Points are deducted because the actual helper scripts are not provided and there is no real CI, test coverage, third-party execution evidence, or independent corroboration; conclusions rely mainly on author documentation, so the score remains at or below 5.