The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and store untrusted audio transcripts verbatim in markdown files, creating a surface for potential injection attacks.
Ingestion points: Audio messages and transcripts are ingested from channels like Telegram or voice memo uploads as described in the 'When to invoke' section of SKILL.md.
Boundary markers: The skill uses block quotes (>) and specific section headers (## User's Words) to delimit user input, which provides some structural separation but lacks explicit 'ignore embedded instructions' warnings for the agent.
Capability inventory: The skill has write access to multiple directories (originals/, concepts/, people/, companies/, ideas/, personal/, voice-notes/) and triggers transcription tools.
Sanitization: The skill explicitly mandates that no sanitization or editing occur ('Never paraphrase. Never clean up. ... transcript section is sacred.'), which ensures that any malicious instructions contained within a voice note are preserved exactly as spoken.
[COMMAND_EXECUTION]: The skill instructions include the use of shell-based utilities for audio manipulation.
Evidence: The 'When to invoke' section of SKILL.md specifies the use of ffmpeg to segment audio files larger than 25MB. This involves the execution of shell commands on user-provided audio files.

voice-note-ingest