The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructions define shell command sequences for site validation. These include using curl to verify local server routes (e.g., http://localhost:4321), rg (ripgrep) for linting anchor links in source files, and sqlite3 for querying the local database to verify that content has been successfully seeded.
[DYNAMIC_EXECUTION]: The helper script scripts/restaurant_site_data.js uses spawnSync to execute the dineway CLI through npx for media uploads. Furthermore, the skill instructions guide the agent to use npx for generating types, seeding content, and deploying the site to the vendor's infrastructure.
[INDIRECT_PROMPT_INJECTION]: The skill possesses an attack surface for indirect prompt injection because it is designed to ingest and process untrusted data from external sources, including Google Place JSON, restaurant reviews, and official website content, and then re-write this information into public site content.
Ingestion points: Untrusted data enters the agent context via places/${placeId}.json files and by crawling URLs extracted from the websiteUri field.
Boundary markers: The instructions mandate the creation of a 'Customer Visibility Gate' and 'Brand Voice & Copy Tone Brief' (documented in .plan/findings.md) intended to guide the agent in filtering out internal instructions or malicious prompts embedded in source data.
Capability inventory: The skill can perform file writes, execute shell commands for validation, and perform network requests to download images and crawl websites.
Sanitization: The skill implements a 'visibility filter' as a procedural step where the agent must internally verify if content is appropriate for a guest before generating code or content files.

dineway-building-restaurant