The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill creator provides templates for new skills that ingest HTML content from external URLs for automated processing, which is a known vector for indirect prompt injection.
Ingestion points: The SKILL.md template explicitly instructs the agent to download HTML from a downloadUrl into a local file (temp/source.html) for further conversion.
Boundary markers: The provided template lacks explicit instructions or delimiters to warn the agent that the downloaded HTML may contain untrusted instructions.
Capability inventory: Skills created with this tool are intended to have access to Bash, Read, Write, and specialized stitch MCP tools, providing a significant capability tier for any injected instructions to exploit.
Sanitization: The scaffolded workflow does not include sanitization, validation, or filtering steps for the ingested external content.
[COMMAND_EXECUTION]: The skill includes a utility script scripts/init_stitch_skill.py that executes file system operations to bootstrap new skill directories.
The script uses pathlib to create directories and write template files based on user-provided skill names.
It includes input validation using a regular expression (stitch-[a-z0-9]+(?:-[a-z0-9]+)*) to mitigate path traversal risks during the initialization process.
[EXTERNAL_DOWNLOADS]: The documentation and templates reference a script (fetch-stitch.sh) designed to download HTML design files from Google Cloud Storage (GCS).
This pattern is documented as a reliable method for fetching assets from a well-known cloud service provider.
The reference targets the vendor's own internal tools (stitch-mcp-get-screen) for consistent asset retrieval.

stitch-skill-creator