The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted input in the form of text, files, or URLs which are then used to generate prompts for subagents.
Ingestion points: Step 1 of SKILL.md identifies the Text Artifact (inline text, local path, or URL) as the primary input.
Boundary markers: The correction brief in references/correction.md wraps content in <text> and <findings> tags and instructs subagents to only modify what the findings concern. SKILL.md Step 6 specifically excludes code blocks from the review process.
Capability inventory: The skill possesses the ability to read local files and external URLs, perform in-place file edits, and execute local Python scripts via the uv tool.
Sanitization: Step 6 of SKILL.md mandates that fenced blocks and inline code be treated as quoted material, preventing instructions within them from affecting the agent's behavior.
[COMMAND_EXECUTION]: The skill executes local scripts to perform its logic.
It runs scripts/invoke.py using uv run, passing raw user input to the backend engine via stdin.
It invokes languages.py from the collection library to resolve editorial resources.
[DYNAMIC_EXECUTION]: The scripts/invoke.py script searches for a vendor-provided engine (kntnt.py) at runtime by scanning a predefined list of common agent skill directories within the user's home directory.
[EXTERNAL_DOWNLOADS]: The skill allows for the retrieval of content from arbitrary URLs provided as input for the purpose of editorial review.

redline