The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests data from the user's screen history (window titles and OCR text) via the local screenpipe API, which acts as an untrusted input source.
Ingestion points: scripts/fetch_window.py retrieves timeline events including window_title and text from the local screenpipe HTTP service.
Boundary markers: The data is interpolated into the LLM prompt in scripts/synthesize.py without strict delimiters, creating a surface where malicious content on a user's screen could attempt to influence the skill's logic.
Capability inventory: The skill possesses file-write and directory-move capabilities (scripts/run.py, scripts/promote.py) used for creating and installing new skills.
Sanitization: scripts/redact.py provides defense-in-depth by scrubbing sensitive patterns (API keys, tokens, emails) from the screen data before it is sent to the LLM.
[DYNAMIC_EXECUTION]: The skill uses dynamic imports and script generation as part of its core functionality.
Evidence: scripts/autoskill.py uses importlib.import_module to load subcommand modules like run at runtime.
Evidence: scripts/run.py generates new SKILL.md files based on LLM output. This is a standard behavior for this skill's purpose as the proposals are stored in a temporary directory for user review before being manually promoted.

autoskill