The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill instructions in SKILL.md and reference.md mandate that if the host is macOS, the agent must follow an external installation reference for the 'ego-browser' tool if it is not found. This involves downloading and executing an unverified installation script at runtime.
[DYNAMIC_EXECUTION]: In the 'Preflight' and 'Workflow' sections of SKILL.md, the agent is instructed to 'Build the smallest probe the project toolchain can run when none exists, then run it.' This encourages the generation and execution of arbitrary code to perform layout measurements at runtime.
[COMMAND_EXECUTION]: The skill workflow relies on executing shell commands, including 'layout-probe' commands, 'command -v ego-browser', and a local validation script 'scripts/check-figma-verify-report.mjs'. It also utilizes JavaScript evaluation (e.g., 'getComputedStyle') to interact with the environment and running applications.
[EXTERNAL_DOWNLOADS]: The skill fetches design data and node trees from Figma (figma.com) and reference exports. While Figma is a well-known service, these external data sources are used to drive the measurement and reconciliation logic, creating an external data dependency.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from external Figma files (node-tree data, reference exports) which are then used to derive 'expected' values and reconciliation tables.
Ingestion points: Figma node-tree data and reference exports fetched via URL as described in SKILL.md (Step 2).
Boundary markers: No specific delimiters or safety warnings are present to distinguish external Figma data from agent instructions.
Capability inventory: Shell command execution, dynamic script generation/execution, file system writes (reports and screenshots), and JS evaluation via browser channels.
Sanitization: The skill lacks explicit sanitization or validation of the fetched Figma data before it is processed by the measurement logic.

figma-pixel-verify