The Agent Skills Directory

[CREDENTIALS_UNSAFE]: The skill documentation explicitly instructs the agent to harvest credentials from sensitive system configuration files rather than following secure input practices.
In Phase 7 of SKILL.md, the instructions tell the agent to "check their environment first" and specifically point to ~/.zshrc and a file named reference_gemini_api_keys.md to locate a "paid-tier EasyChamp Gemini key."
This promotes the unauthorized collection of secrets from shell histories and environment configuration files.
[OBFUSCATION]: The skill uses zero-width characters (U+200B) to obscure technical keywords in both documentation and script comments, a technique often used to bypass string-matching security filters.
Multiple instances of the Zero-Width Space (U+200B) were found within the word "concat" (e.g., concat) in SKILL.md and scripts/render_master.py.
[COMMAND_EXECUTION]: The skill includes Python scripts that execute shell commands via the subprocess module, which are used for video and audio processing.
scripts/render_master.py and scripts/scrub_beeps.py both utilize subprocess.run() to invoke FFmpeg.
While these are functional for the skill's purpose, the execution of shell commands using parameters derived from external transcription data increases the risk of command injection.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted transcript data and requires the agent to "Read the full transcript," providing a vector for indirect injection attacks.
Ingestion Points: Transcript data is ingested from .work/source.json (Whisper output) in Phase 2.
Boundary Markers: The skill lacks explicit delimiters or instructions to treat the ingested transcript content as untrusted.
Capability Inventory: The agent possesses the capability to execute shell commands (ffmpeg) and perform network operations (curl).
Sanitization: No filtering, escaping, or validation is performed on the transcript content before it is processed by the agent.

hyperframes-best-practices