The Agent Skills Directory

[SAFE]: The skill implements proactive data protection and secure credential management.\n
Redaction layer: scripts/redact.py uses regex patterns to scrub private keys, API tokens, JWTs, and PII from OCR text and window titles.\n
Secret management: Uses environment variables (SCREENPIPE_TOKEN, ANTHROPIC_API_KEY) and explicitly avoids hardcoded credentials in code or config templates.\n- [DATA_EXFILTRATION]: Access to sensitive screen history is managed locally.\n
Data source: Fetches screen OCR and UI data from a local screenpipe daemon (http://localhost:3030) over loopback.\n
Exfiltration control: Defaults to local LLM endpoints (LM Studio) to keep data on-host; external network calls are opt-in and authenticated.\n- [PROMPT_INJECTION]: Indirect injection surface exists in the processing of untrusted screen content.\n
Ingestion points: fetch_window.py reads user-controlled screen text and window titles from the screenpipe daemon.\n
Boundary markers: Lacks explicit 'ignore embedded instructions' delimiters in the synthesize.py prompt template when interpolating window titles.\n
Capability inventory: Writes SKILL.md instruction files to the local filesystem (run.py) and moves them to the skills directory (promote.py).\n
Sanitization: Lacks specific prompt injection sanitization, relying instead on user review of generated drafts before promotion.

autoskill