The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process external legal materials, including samples, SOPs, and user-provided case summaries, which presents an attack surface for indirect prompt injection. This is mitigated by explicit instructions to recognize and prompt for the de-identification of sensitive information (as detailed in references/alignment-strategies.md) and the requirement to label all unverified legal sources as 'unverified' in the generated output. The skill output is a structured markdown brief ('Legal Skill Brief v1') which limits the ability of untrusted data to execute commands, as the skill itself does not perform shell or network operations.
[COMMAND_EXECUTION]: Evaluation and compatibility documentation (e.g., evals/case-bundle-260728/b6-compatibility-analysis/README.md) references the execution of local scripts for validation purposes, such as python -m scripts.quick_validate and python -m scripts.run_eval. These commands are part of the developer workflow and testing harness rather than the skill's operational logic, and they target local files rather than remote or untrusted code sources.

legal-skill-alignment