The Agent Skills Directory

[DATA_EXFILTRATION]: The skill processes highly sensitive local screen OCR data provided by the screenpipe daemon. To mitigate risks, it implements several security layers:
Redaction: A specialized script (scripts/redact.py) uses extensive regex patterns to scrub PII and secrets (API keys, JWTs, private keys, emails) from the OCR data before it is processed by any LLM.
Local-First Default: The configuration defaults to a local LM Studio instance (localhost:1234), ensuring sensitive data does not leave the user's machine.
Authorized Access: Requests to the local screenpipe API and cloud LLM providers use environment variables for authentication, avoiding hardcoded credentials.
[PROMPT_INJECTION]: As the skill ingests untrusted text from the user's screen (e.g., from a website or document), it is theoretically susceptible to indirect prompt injection. A malicious string displayed on the screen could attempt to influence the skill generation logic. This risk is addressed by the skill's architecture, which requires explicit user triggers for each analysis and a manual review/promotion step for any drafted skills.
[COMMAND_EXECUTION]: The skill is designed to generate new SKILL.md files which may include instructions for the Bash tool. While the generation of executable content is a primary function of the skill, security is maintained through a two-stage process: proposals are written to a temporary directory (~/.autoskill/proposed/), and the user must manually review and execute a promote command to integrate them into the active skill library.

autoskill