The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill provides instructions for ingesting untrusted data, such as transcripts from webinars or podcasts, to create new agent skills. This creates an indirect prompt injection surface. However, the skill includes robust defensive measures:
Ingestion points: Untrusted data enters the context through transcripts and source registers as described in references/content-to-skill.md and references/content-to-skill-corpus-design.md.
Boundary markers: The instructions mandate separating editorial synthesis from source material and explicitly labeling omissions or gaps in the data.
Capability inventory: No risky capabilities such as subprocess execution, file writing, or network operations are present in the skill.
Sanitization: The skill contains explicit instructions to treat all source material as data and specifically identifies malicious patterns (e.g., 'ignore previous instructions') as content that must be ignored by the agent.
[PROMPT_INJECTION]: Static analysis identified the phrase 'ignore previous instructions' in references/content-to-skill-corpus-design.md. Review confirms this is a defensive instruction and a worked example used to teach the agent how to resist prompt injection in source material, rather than a malicious attempt to override behavior.

lead-magnets