The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted user input ('brain dumps') which are used as keywords for filesystem searches via the provided scripts. While this creates an attack surface for influencing the agent's workspace discovery, the risk is negligible as the skill lacks network exfiltration capabilities and mandates explicit user approval for any proposed actions.
Ingestion points: Processes unstructured user text through explicit triggers (e.g., 'brain dump') and implicit signals defined in SKILL.md.
Boundary markers: The skill relies on natural language instructions for organization but does not implement strict sanitization or delimiters for the user-supplied content before it is processed by internal scripts.
Capability inventory: Includes scripts/workspace_inventory.py which provides the agent with the ability to scan the local filesystem, glob filenames, and grep file contents for keyword matches.
Sanitization: Input text is categorized using regex-based heuristics in scripts/dump_classifier.py and analyzed for keyword frequency in scripts/complexity_estimator.py.
[COMMAND_EXECUTION]: The skill provides and utilizes local Python scripts (scripts/workspace_inventory.py, scripts/dump_classifier.py, scripts/complexity_estimator.py) for data processing. These scripts are implemented using the Python standard library, do not perform unsafe shell execution, and are restricted to local processing and filesystem discovery as intended by the skill's functionality.

capture