The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill fetches external instructions and supporting references from third-party GitHub repositories (emilkowalski/skills and jakubkrehel/skills). While the instructions explicitly state not to run downloaded scripts or start subagents, the agent is directed to follow the principles and guidance contained within these unverified remote files.
[COMMAND_EXECUTION]: The skill executes various local commands including project build, lint, and test scripts. It also uses browser automation tools like chrome-devtools-cli for verification and tools like hitch to inspect running servers and environment files.
[INDIRECT_PROMPT_INJECTION]: The skill possesses a surface for indirect prompt injection by ingesting and processing untrusted data from multiple sources.
Ingestion points: Processes local design documents, project codebases, and external markdown guidance fetched from GitHub.
Boundary markers: Includes instructions to "name the missing inputs" instead of inventing them and to prioritize project conventions over external guidance, though no explicit sanitization of input text is defined.
Capability inventory: The agent can execute shell commands (via project scripts), interact with browsers, and read environment files.
Sanitization: There is no evidence of specific escaping or validation for content parsed from design artifacts or external guidance files.

webcraft-build