The Agent Skills Directory

[PROMPT_INJECTION]: The skill describes an architecture for 'Indirect Prompt Injection' where models fill Markdown checklists that are then parsed by scripts to drive deterministic decisions. \n
Ingestion points: Markdown checklists filled by LLM models (defined in SKILL.md).\n
Boundary markers: Relies on deterministic parsing of checkbox patterns (- [x], - [ ]) but lacks explicit sanitization of model-generated text within those items.\n
Capability inventory: The system executes shell commands (agent-harness), performs file I/O for reports, and mentions use of codex exec for code execution.\n
Sanitization: Instructions suggest parsing checkboxes deterministically but do not define validation for prose or evidence fields populated by the model.\n- [COMMAND_EXECUTION]: The skill provides shell script templates for implementing 'Stop Hooks' that invoke CLI tools and Python interpreters. \n
Evidence: SKILL.md contains an example script that executes agent-harness run-checks and pipes a Python script via heredoc (`python3
<<'PY'`) to process output.\n- [COMMAND_EXECUTION]: The skill provides templates for executing Python logic via shell heredocs to process harness reports and allows loading code from local vendor paths. \n
Evidence: The 'Minimal Stop hook pattern' in SKILL.md uses dynamic Python execution to parse JSON reports and optionally modifies PYTHONPATH to include local sibling directories.

harness-check-designer