Legal Research

legal-research-py

A Claude Code skill for autonomous, multi-jurisdiction legal research. It produces verifiable, citation-backed legal reports by orchestrating a sequence of fresh-context subagents — each handling one phase of the research workflow in its own context window.

Install

npx skills add Flosters/legal-research2

Or clone manually and place the directory in `~/.claude/skills/`.

---

What it does

Given a legal question, the skill:

1. Extracts the research scope — jurisdiction, area of law, legal posture, language
2. Designs a verifiable checklist of legal nodes (issues) and N targeted research queries
3. Creates a dedicated NotebookLM notebook for the research session
4. Runs N deep-research queries in parallel (one sub-subagent per query, each with its own temp notebook — eliminates the research status race condition)
5. Curates sources against a jurisdiction-aware priority ladder, deduplicates, and builds an Evidence Registry
6. Imports primary sources (T1) with crawlability checks and fallback URL resolution
7. Spot-checks every T1 source for actual queryability in parallel (one sub-subagent per source)
8. Runs an IRAC-structured analysis sequence via NotebookLM chat
9. Verifies every citation against the notebook — marks each as ✓ Verified, ~ Paraphrase — Consistent, or [SECONDARY ONLY]
10. Cross-examines weak propositions via adversarial prompting
11. Assembles and writes an HTML report

All phases are stateless across subagents — state passes through a schema-validated `state.json` file in the workspace directory.

---

Workflow

User query
    │
    ▼
Phase 1 — Scope extraction + checklist + query design  (orchestrator, inline)
Phase 2 — Notebook creation                            (orchestrator, inline)
    │
    ▼  state.json written → autonomous from here
    │
Phase 3   — Parallel deep research (Subagent A)
              └─ N query runners dispatched simultaneously (sub-subagents)
Phase 3.5 — Source cu