The Agent Skills Directory

[COMMAND_EXECUTION]: The skill performs local command execution as part of its validation workflow. It invokes python3 to run a quick_validate.py script and uses the qodercli tool to validate plugins and perform forward-tests.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection because its core workflow involves ingesting and analyzing untrusted data from external repositories, evaluation corpora, and workflow chains.
Ingestion points: The agent is instructed to inspect source chains in external repositories, including entry SKILL.md files, scripts, validators, and manifests (defined in SKILL.md Workflow Step 2).
Boundary markers: The instructions include a 'Prompt contract' in references/bootstrap-patterns.md that attempts to constrain the model output to a compact JSON object and explicitly forbids writing files or emitting tool calls during the inspection phase.
Capability inventory: The environment allows for file system inspection, git status checks, and command execution via python3 and qodercli.
Sanitization: The skill provides guidelines to reject output that includes 'pseudo tool calls', 'shell probes', or instructions that copy source product policies as universal rules.

harness-skill-creator