The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted HTML and CSS content from user-provided sources to perform design analysis and conversion. This content is processed by the agent to infer design intent, creating a surface where hidden instructions in the source code could influence agent behavior.
Ingestion points: The skill reads content via direct pastes, local file access, URL fetching, and DOM rendering (documented in phases/phase-1-forensics.md).
Boundary markers: There are no instructions to use delimiters or "ignore embedded instructions" warnings when processing the untrusted input.
Capability inventory: The skill utilizes shell execution for auditing (grep, wc, tr), file writing for outputting converted code, and external tool invocation via Figma and System MCP servers.
Sanitization: No sanitization or filtering of the input HTML/CSS content is specified before analysis or interpolation.
[EXTERNAL_DOWNLOADS]: The skill fetches external content from user-provided URLs for forensics and references remote assets from well-known services (e.g., Google Fonts, Tailwind CDN, unpkg, jsDelivr) within the generated code artifacts.
Evidence: phases/phase-1-forensics.md (fetching URLs), intake/SYSTEM-INTAKE.md (CDN URLs for tokens and components), and reference/THE-TELLS.md (Tailwind Play CDN).
[COMMAND_EXECUTION]: The skill uses standard shell utilities to perform deterministic counts and static analysis on the source code to provide empirical data for its "autopsy" and "receipt" reports.
Evidence: phases/phase-1-forensics.md and phases/phase-4-proof.md contain commands using grep, sort, uniq, wc, and tr to count hardcoded values and CSS properties.

vibe-killer