gfodor/legal-skills

What it solves

This project aims to displace expensive legal fees associated with U.S. utility patent work. It replaces the manual process of hiring a patent boutique for drafting, auditing, and filing preparation with a set of executable AI agent skills and deterministic Python scripts, allowing inventors and founders to perform high-level patent audits and simulations themselves.

How it works

The system operates as a set of "agent skills" (Markdown-based protocols) that can be integrated into AI agents like Claude Code or Grok. It uses a three-stage pipeline:

1. patent-audit: A pre-filing audit using a 316-item checklist based on U.S. statutes and USPTO practice. It employs a deterministic Python pre-pass for mechanical checks (like date arithmetic and numeral consistency) and a fan-out of specialist agents to assess legal and technical gaps. It includes an adversarial verification phase where findings are re-tried by fresh verifiers.
2. patent-examine: A simulated USPTO prosecution loop. It spawns independent simulated examiners who issue mock Office Actions. The goal is to reach a state where two different examiners independently clear the same version of the application.
3. patent-workaround: Adversarial testing where subagents attempt to design around the claims to identify weaknesses and ensure the patent is robust against competitors.

Who it’s for

• Inventors, founders, and in-house engineers who want to audit their patent drafts or simulate the examination process without immediately opening a legal matter.
• Law firms looking to make their drafting and filing process faster and more cost-effective.

Highlights

• 316-item Audit Checklist: A comprehensive, primary-source based audit covering everything from statutory bars to drawing formalities.
• Deterministic Guardrails: Uses Python for tasks LLMs struggle with, such as date calculations, claim dependency graphs, and numeral reconciliation.
• Multi-Agent Adversarial Loop: Employs red-team examiners and design-around agents to stress-test claims.
• State-Locked Prosecution: Maintains a strict separation between ground truth, proposed candidates, and the prosecution record to ensure version control during simulated examination.