The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill defines the 'Nen Court' protocol, which involves external validator scripts that process potentially untrusted data to enforce project constraints.
Ingestion points: The validator scripts (e.g., markdown_wrap.py, iron_law.py, proof_of_work.py) ingest data via stdin. This data includes user-provided markdown files, git commit logs, and agent-generated text.
Boundary markers: The protocol utilizes specific exit codes (0, 1, 2) and JSON-formatted input/output to define boundaries, though it does not explicitly detail sanitization methods for the content within the JSON objects.
Capability inventory: The system allows the execution of local Python scripts to audit project state. A violation (exit code 1) is designed to block mission progress unless overridden by a user.
Sanitization: The instructions imply structured data handling (JSON), but the underlying scripts must handle potentially malicious instructions embedded in the analyzed text (e.g., within commit messages or markdown files) to prevent logic subversion.
[COMMAND_EXECUTION]: The skill provides explicit instructions and examples for executing shell commands to run local validation scripts.
Evidence: The 'Validator Invocation' section documents command patterns such as python plugins/imbue/validators/markdown_wrap.py and python plugins/imbue/validators/iron_law.py using piped JSON input.

vow-enforcement