The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill has a large ingestion surface for untrusted data as it reads project documents (PRDs, specs, READMEs) and fetches external URLs in 'study' and 'audit' modes.
Ingestion points: SKILL.md, references/verbs/study.md, and references/verbs/audit.md where external content is parsed.
Boundary markers: Present. The skill explicitly instructs the agent that project documents are 'data, not instructions' and to ignore any directives embedded in fetched HTML or CSS.
Capability inventory: The skill can fetch network assets (icons, photos) and write to local project directories (design/assets/). It does not have arbitrary command execution capabilities.
Sanitization: Content is treated as inert design data for attribute extraction.
[DYNAMIC_EXECUTION]: The script scripts/generate_palette.py uses __import__("os") to dynamically resolve the current file's directory for path manipulation. This is used for internal module resolution and does not process external input.
[OBFUSCATION]: The file scripts/test_security.py contains a Base64-encoded string (PHNjcmlwdD4=) which decodes to a <script> tag. This is a non-functional snippet used exclusively as a test case to verify that the skill's network guard correctly blocks XSS-related data URIs.
[EXTERNAL_DOWNLOADS]: The skill fetches design assets from several established services.
Fetches icons from Iconify (api.iconify.design).
Fetches photography from Openverse (api.openverse.org) and Pixabay (pixabay.com).
Downloads specific JavaScript libraries (GSAP, Motion) from CDNs (cdnjs.cloudflare.com, cdn.jsdelivr.net).
These operations are performed via scripts/_netguard.py which enforces a strict scheme and host allowlist.

tastemaker