The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process external, untrusted data such as meeting transcripts and help documentation, which creates a potential surface for indirect prompt injection.
Ingestion points: The skill includes functions in SKILL.md and examples.md (e.g., load_vtt, load_livekit_transcript, load_recall_transcript) that read external file content into the agent's context.
Boundary markers: The implementation uses dspy.Signature classes to provide structured input/output fields, which helps separate context from questions, though it lacks explicit adversarial instruction filtering.
Capability inventory: The skill utilizes dspy for LLM calls and local file reading via Python's open() function. No file-writing or network exfiltration capabilities are present in the provided code.
Sanitization: While it does not sanitize for malicious instructions, the skill specifically implements 'faithfulness verification' as a security control to prevent the AI from generating unsupported claims.
[EXTERNAL_DOWNLOADS]: The skill requires the installation of external Python libraries dspy and langfuse. It also suggests downloading additional modular components from the same author's repository using the npx skills add command.
[COMMAND_EXECUTION]: The documentation provides command-line instructions for the user to manage and install related skills using a platform-specific skill management utility (npx skills).

ai-stopping-hallucinations