The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill describes a process for converting external content such as webinar transcripts or podcast archives into new agent skills. It identifies this as a potential attack surface and provides clear instructions to treat such data as untrusted. * Ingestion points: Processed content libraries mentioned in references/content-to-skill.md. * Boundary markers: Explicit instructions to treat source material as data and ignore embedded commands appear in references/content-to-skill-corpus-design.md. * Capability inventory: No dangerous capabilities like network or file-write tools are invoked in the provided files. * Sanitization: Guidelines for separating source claims from agent instructions are provided to prevent command confusion.
[SAFE]: The inclusion of security-focused evaluation cases in evals/evals.json demonstrates a defensive design, testing for resilience against data-embedded prompt injections and exfiltration attempts.

lead-magnets