The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes user-provided inputs such as 'site requirement descriptions', 'requirement briefs', and 'existing site content' to populate the SITE.md file. This creates an attack surface where malicious instructions embedded in these external sources could influence the agent's behavior.
Ingestion points: Identified in SKILL.md (Step 1) and references/workflow.md (Action Plan), where the agent is instructed to locate and read project description documents or text inputs.
Boundary markers: The skill uses a structured template (six specific sections: Identity, Visual, Architecture, Sitemap, Backlog, and Creative Boundaries) which helps limit the influence of the input data to specific fields.
Capability inventory: The skill is authorized to create/write files (specifically .stitch/SITE.md) and read existing project files.
Sanitization: While the instructions explicitly mandate removing PII and credentials (like customer names, phones, and session cookies), there is no specific sanitization mentioned for bypassing instructions contained within the analyzed text.
[EXTERNAL_DOWNLOADS]: The skill references an external repository for implementation details and upstream snapshots.
Reference: The SKILL.md file links to https://github.com/google-labs-code/stitch-skills/tree/0337446dadde6f8c94210444e2aa9d546126480f. This is a reference to an official repository belonging to a well-known organization.

stitch-site-md