The Agent Skills Directory

[CREDENTIALS_UNSAFE]: The file evals/fixtures/dictionary-flawed/src/search.tsx contains a hardcoded string API_KEY = "dict_live_sk_8f3a9b2c1d4e5f6a7b8c9d0e".
Evidence: The source code explicitly labels this as a "deliberately planted hardcoded secret" intended for testing the skill's store-review detection capabilities. As a dummy value in a pedagogical fixture, it does not pose a real-world security risk.
[INDIRECT_PROMPT_INJECTION]: The skill documents the creation of AI Tools which ingest untrusted data from external services and user inputs.
Ingestion points: External REST APIs fetched via useFetch (e.g., src/search.tsx) and arguments passed to AI tools in src/tools/.
Boundary markers: The skill instructs users to define tool input schemas using the Input type and use JSDoc to guide the LLM's parameter selection.
Capability inventory: Extensions can perform network operations, local file system access (via useExec or useSQL), and state mutations.
Sanitization: The skill mandates the use of a confirmation export for any tools performing destructive or mutating actions, ensuring user oversight for high-risk operations.

raycast-extensions