The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to analyze external websites and their Lighthouse reports, which creates a potential surface for indirect prompt injection if those sources contain malicious instructions.
Ingestion points: Untrusted content is ingested from measured URLs and Lighthouse JSON reports (SKILL.md, Phase 2).
Boundary markers: The skill includes a specific safeguard, Guardrail 9, which instructs the agent to treat page content as evidence only and never as instructions.
Capability inventory: The skill can execute shell commands via Lighthouse, write findings to the repository, create GitHub issues, and hand off tasks to an orchestrator.
Sanitization: The skill relies on natural language instructions to separate external data from the agent's control flow.
[EXTERNAL_DOWNLOADS]: The skill utilizes external packages and tools for measurement and asset optimization.
Evidence: Instructions in references/lighthouse-verification.md include npx lighthouse (a well-known tool from Google) and pip install fonttools brotli for font processing.
[DYNAMIC_EXECUTION]: The skill employs runtime scripts to parse audit results and perform font manipulations.
Evidence: Uses a node -e one-liner to interrogate Lighthouse JSON files and a Python script using the fontTools library for font instancing and subsetting in references/lighthouse-verification.md.
[COMMAND_EXECUTION]: The skill triggers shell commands to perform audits.
Evidence: Executes npx lighthouse with various flags to generate performance reports.

web-design-guidelines