The Agent Skills Directory

[COMMAND_EXECUTION]: The skill performs shell command execution using a utility wrapper for subprocess.run found in tools/stdio_utf8.py. This is used to invoke the skill's own suite of Python auditing and rendering tools as part of the patent drafting workflow.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process untrusted patent disclosure materials (Markdown and YAML files) provided by the user in a specified directory.
Ingestion points: Files are read from the user-specified --case-dir by multiple tools including material_gate.py, audit_claims.py, and check_numeral_register.py.
Boundary markers: The prompts establish strict formatting rules for the output but do not explicitly define boundary delimiters or "ignore instructions" markers for the ingested content to separate data from instructions.
Capability inventory: The skill possesses capabilities for shell command execution (via Python script invocation), file system writes (generating Word documents and images), and headless browser execution (via Playwright).
Sanitization: The tools apply structural validation using regex and schemas, and the SVG rendering logic utilizes HTML escaping to prevent injection within generated assets.
[EXTERNAL_DOWNLOADS]: The skill scripts reference and utilize several external Python libraries including playwright, PyYAML, python-docx, and latex2mathml. These are standard, well-known libraries required for the skill's documented functionality of document parsing, rendering, and conversion.

patent-application