The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes high-stakes legal deliverables (such as opinions, M&A memos, and litigation documents) which represent a significant ingestion surface for untrusted data. Maliciously crafted content or instructions embedded within these documents could attempt to bias the adversarial agents (e.g., instructing the 'Attacker' to overlook specific flaws or influencing the 'Synthesizer' to favor a specific conclusion).
Ingestion points: External legal documents (deliverables) provided by the user for review (SKILL.md).
Boundary markers: The skill employs an 'Isolated' reviewer mode that restricts specific agents to reading only the logic of a draft without access to the source content (SKILL.md). It also recommends pseudonymization via the let-it-be skill.
Capability inventory: The skill utilizes Read and Write tools to process documents, retrieve case law via companion skills, and generate multiple draft versions and audit reports (SKILL.md Frontmatter).
Sanitization: The workflow includes a mechanical 'Verifier' role with a 10-point checklist and citation-grounding requirements to validate claims against legal sources, reducing the impact of hallucinated or injected misinformation.
[COMMAND_EXECUTION]: The skill uses Read and Write tools to facilitate its automated debate and revision loop. It manages file versioning (draft_v1, draft_v2) and writes audit transcripts to the local file system. While these are legitimate functions of a legal review tool, they represent the capability to modify local project files based on agent-generated content.

adversarial-legal-review-pl