The Agent Skills Directory

[DYNAMIC_EXECUTION]: The script scripts/invoke.py dynamically identifies an execution engine (kntnt.py) by searching through a pre-defined list of global skill directories, such as ~/.claude/skills, ~/.config/goose/skills, and ~/.cursor/skills. Once a valid manager directory is found, it executes the engine using subprocess.run. This runtime resolution of executable paths from potentially user-writable locations is a security concern.
[COMMAND_EXECUTION]: The skill invokes local Python scripts using uv run. Specifically, SKILL.md instructs the agent to run scripts/invoke.py, which in turn spawns a subprocess to run the discovered kntnt engine. The skill also executes scripts/languages.py from an external library path.
[EXTERNAL_DOWNLOADS]: The skill is designed to accept a URL as an input operand. The agent will fetch content from the provided URL to perform proofreading tasks.
[INDIRECT_PROMPT_INJECTION]: The skill has a significant attack surface for indirect prompt injection.
Ingestion points: The skill ingests untrusted data from inline text, local files, and remote URLs (identified in SKILL.md and help.md).
Boundary markers: While the skill instructs the agent to treat code blocks as "quoted material" and not modify them, it lacks explicit delimiters or "ignore instructions" markers for the primary text artifact being processed.
Capability inventory: The skill possesses capabilities for subprocess execution (scripts/invoke.py), file writing (via the --in-place option), and network reading (via the URL operand).
Sanitization: No explicit sanitization or validation of the ingested content is performed before processing.

proofread