The Agent Skills Directory

[PROMPT_INJECTION]: The skill handles untrusted data by reading release content from a database (r.content) which originates from external sources like changelogs and blogs. This presents an indirect prompt injection surface where external content could influence the AI's generation process.
Ingestion points: Data is pulled from the releases table using wrangler d1 execute in Step 1.
Boundary markers: The prompt template for sub-agents instructs them to use specific XML tags and discard irrelevant ones, providing some structural delimiters.
Capability inventory: The skill can execute database queries via wrangler, perform CLI updates via releases admin, and send network requests via curl.
Sanitization: The skill utilizes the isEmptyContent function to skip processing for boilerplate or empty content and parseReleaseContent to validate AI output structure.
[COMMAND_EXECUTION]: The skill relies on several local command-line tools to perform its tasks.
It executes bunx wrangler d1 execute to perform SQL queries against a remote database.
It uses a custom CLI tool releases admin release update to write data to the local environment.
[DATA_EXFILTRATION]: The skill performs network operations to persist data to an external service.
It uses curl to send PATCH requests to $RELEASES_API_URL.
Authentication is handled via RELEASES_API_KEY, which is expected to be provided as an environment variable.

generating-release-content