The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses a platform-specific Workflow() function to execute local TypeScript scripts located in the .claude/workflows/ directory. This is used for both single-source backfills (backfill-source.ts) and bulk operations (backfill-sweep.ts).
[EXTERNAL_DOWNLOADS]: The skill is designed to fetch content from external URLs provided in the source argument (e.g., acme/changelog) to perform extraction and ingestion of releases.
[PROMPT_INJECTION]: The skill exhibits an indirect prompt injection surface as it ingests and processes untrusted data from external changelog sources.
Ingestion points: Data enters the agent context via the source argument in Workflow calls, which triggers a fetch and extract loop.
Boundary markers: The skill implements a 'Preflight gate' that checks for ai-input=no or ai-train=no signals to respect publisher opt-outs, failing closed if found.
Capability inventory: The environment allows for script execution via Workflow, filesystem writes to ~/.releases/work/, and network requests to internal vendor APIs.
Sanitization: Content is processed through LLM-based extraction (Sonnet/Haiku) and includes an 'altitude check' to detect anomalies in record counts relative to the date span.
[DATA_EXFILTRATION]: The skill references the use of RELEASES_API_* environment variables and makes authenticated network requests to internal API endpoints (e.g., POST /v1/workflows/backfill-media) for administrative tasks like media mirroring and re-extraction. These are consistent with the vendor's documented infrastructure.

backfilling-sources