The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests data from external websites (Awwwards, Mobbin) to generate a 'Reference Board' stored in .sumi/refs/board.json. This board serves as a contract that other commands, such as /style and /grade, consume to guide their behavior and output.
Ingestion points: External site metadata and rendered content (DNA) fetched via scripts/awwwards.mjs and scripts/capture.mjs (SKILL.md).
Boundary markers: The instructions do not define specific sanitization protocols or boundary markers for the ingested data before it is incorporated into the board.json contract.
Capability inventory: The skill executes shell commands, writes to the local filesystem, and performs network operations via browser automation (capture.mjs).
Sanitization: There is no mention of filtering or sanitizing instructions that could be embedded in the design references (e.g., in text content or metadata), which could potentially influence the agent's behavior.
[COMMAND_EXECUTION]: The skill instructs the agent to execute shell commands to resolve the plugin's path and run JavaScript files using Node.js.
Evidence: SKILL.md contains a shell block that uses find to scan ~/.claude/plugins and node to execute internal scripts like scripts/awwwards.mjs and scripts/capture.mjs.
[EXTERNAL_DOWNLOADS]: The skill utilizes browser automation to visit and capture data from arbitrary external URLs.
Evidence: The capture.mjs script uses puppeteer-core to visit live websites and extract metrics, as described in SKILL.md and references/dna-schema.md.

reference-intelligence