The Agent Skills Directory

[DATA_EXFILTRATION]: The skill accesses sensitive screen OCR data through the local screenpipe service. It implements a redaction engine in scripts/redact.py that scrubs common secrets (e.g., AWS, GitHub, and OpenAI keys) and PII (emails, phone numbers) before data is processed by local or optional cloud-based LLM backends.
[PROMPT_INJECTION]: The skill processes untrusted screen content as input for skill synthesis, creating a surface for indirect prompt injection. This risk is mitigated through structured prompt templates and a mandatory manual review and 'promote' workflow for all generated drafts.
Ingestion points: scripts/fetch_window.py retrieves OCR and window title data from the local screenpipe API.
Boundary markers: Synthesis prompts in scripts/synthesize.py use clear headers to separate cluster data from instructions.
Capability inventory: The skill can write generated markdown files to the local file system and move directories between the proposal and repository folders.
Sanitization: Regex-based redaction is applied to all ingested text in scripts/run.py before the data is passed to the LLM backend.
[EXTERNAL_DOWNLOADS]: The skill utilizes the sentence-transformers library, which downloads an embedding model from Hugging Face during the first execution to facilitate local semantic matching against existing skills.

autoskill