The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill downloads the 'sentence-transformers/all-MiniLM-L6-v2' embedding model from Hugging Face during its first run. This is a standard operation for local text processing and originates from a well-known service.
[DATA_EXFILTRATION]: While the skill processes sensitive OCR data from the user's screen, it implements multiple safeguards. It uses a loopback connection (localhost) to communicate with the 'screenpipe' daemon and defaults to a local LLM backend. A dedicated redaction script ('scripts/redact.py') scrubs PII and secrets (AWS keys, tokens, emails, etc.) using regex before sending any summaries to the LLM backend.
[COMMAND_EXECUTION]: The skill uses the 'Bash' tool as specified in its manifest to facilitate the promotion of skills and the running of the 'doctor' diagnostic tool. These executions are user-triggered and operate on locally generated paths.
[PROMPT_INJECTION]: The skill has a surface for indirect prompt injection via OCR window titles processed by the 'synthesize.py' script. However, the impact is mitigated by the structured JSON response format expected from the LLM and the fact that the output consists of drafts for human review rather than immediate executable code.
[REMOTE_CODE_EXECUTION]: The skill does not perform any remote code execution. All logic is contained within local Python scripts, and any generated skill drafts are written to a local staging directory for user inspection.

autoskill