The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill templates demonstrate patterns that ingest untrusted data into the agent's context, creating a surface for potential indirect prompt injection attacks.
Ingestion points: User prompts in templates/ai-text-generation.ts, external image URLs in templates/ai-vision-models.ts which are fetched and processed, and metadata retrieved from Vectorize similarity searches in templates/ai-embeddings-rag.ts.
Boundary markers: The templates generally lack strong delimitation (such as XML tags or explicit instruction markers) between system-provided context and untrusted data.
Capability inventory: The skill utilizes env.AI.run for model inference, env.VECTORIZE.upsert for data indexing, env.BUCKET.put for storage operations, and the fetch API for external network requests.
Sanitization: While references/best-practices.md mentions basic input filtering, the core functional templates interpolate raw external data directly into model prompts.
[EXTERNAL_DOWNLOADS]: The skill references several external dependencies required for the Cloudflare Workers environment.
Packages: hono, zod, workers-ai-provider, ai, and @cloudflare/workers-types.
Status: These are well-known, community-standard packages for building and typed development within the Cloudflare Workers ecosystem.

cloudflare-workers-ai