v0.1.0 to v0.1.1

---

name: case-law-research

description: >

---

# Case-Law Research (bundled MCP servers)

> **NOT LEGAL ADVICE.** This skill retrieves and organizes legal sources as a research and

> drafting aid. It does not evaluate whether an authority is good law on the user's facts and

> does not create an attorney-client relationship. Verify everything against the primary source

> before relying on or filing anything.

## Purpose

The marketplace deliberately snapshots **rules** (statutes, regulations, court rules) verbatim

in each plugin's references corpus, but **case law is never snapshotted** — it is too large and

moves too fast. This skill is the live-lookup layer. This plugin bundles two free remote MCP

servers (declared in [`../../.mcp.json`](../../.mcp.json)), and because every state plugin

depends on this plugin, they are connected anywhere any claude-legal plugin is installed:

| Server | Use it for |

|---|---|

| **CourtListener** (`https://mcp.courtlistener.com/`, Free Law Project) | U.S. case law (federal + all 50 states), RECAP/PACER federal dockets and filings, judges, oral arguments, citation verification |

| **Legal Data Hunter** (`https://legaldatahunter.com/mcp`) | Multi-jurisdictional research — statutes, regulations, case law, and doctrine across 100+ countries; jurisdictions or source types the snapshotted corpora and CourtListener don't cover |

**The cardinal rule: never cite a case from memory.** A case name + reporter cite produced

from the model's weights is a hallucination risk; courts sanction filers for it. Every cited

case must be retrieved through one of these servers (or a canonical fallback) and its

relevant language read before it goes into a draft.

## Research order of operations

1. **Snapshot first.** If the question is about a statute, regulation, or court rule already

mirrored in an installed plugin's `references/` corpus (RCW, ORS, CCP, C.R.S., CPLR,

Tenn. Code Ann., MCL, A.R.S., R.C., the federal corpora here, etc.), cite the verbatim

snapshot — that's what it's for.

2. **CourtListener for U.S. case law and dockets.** Anything decided by a U.S. court.

3. **Legal Data Hunter for everything else** — non-U.S. law, jurisdictions not snapshotted,

doctrine/commentary, or cross-border comparison.

4. **Canonical fallbacks** if a server is unavailable (see "If the tools are missing" below

and the installed state plugin's `legal-data-apis.md`).

## CourtListener playbook

The server exposes search plus full REST-endpoint access. The high-value moves:

- **`search`** — the entry point. Collections: **opinions** (case law), **RECAP**

(federal dockets/filings from PACER), **judges**, and **oral arguments**. Filter by court,

date, and citation. Court identifiers are CourtListener court IDs (`scotus`, `ca1`–`ca11`,

`cadc`, `cafc`, district courts like `wawd`, state courts like `wash` / `washctapp` /

`cal` / `ny` / `ohio`); when unsure, discover them via the courts endpoint or the server's

choices tool rather than guessing.

- **`read_document` / `search_document`** — read an opinion's or filing's text (paginated,

greppable). Use these to quote-check; do not paste reporter quotes from memory.

- **Citation tools** — extract citations from a draft and look them up in bulk to confirm

each resolves to a real case; use the citation network (what cites this case) as a

first-pass treatment check.

- **Field discipline** — responses can be enormous; always restrict to the fields you need.

Don't fetch PDFs; use the document-reading tools.

- **Linking** — cite retrieved cases with their `courtlistener.com` URL alongside the

reporter cite so a human can verify in one click.

**Treatment caution:** CourtListener's citation network shows *who cites a case*, not

editorial negative-treatment analysis. It is **not** Shepard's or KeyCite. If a case is

load-bearing, say so and recommend the user confirm treatment in a citator before filing.

## Legal Data Hunter playbook

- **Discover before searching** — list available countries/sources/filters first when

working outside familiar territory, then run hybrid semantic + keyword `search` scoped by

country, source, and court level.

- **`get_document`** retrieves the full text; **`resolve_reference`** turns a bare citation

into its document. Quote only from retrieved text.

- Best fits: foreign and comparative law, statutes/regulations of jurisdictions the

marketplace doesn't snapshot, doctrine, and multi-jurisdiction surveys (e.g., "how do

other states treat X").

- Cite **inline**, in the sentence where the claim appears, with the document link.

## Output discipline

- Every legal claim in the research output carries an inline citation to a retrieved

document (CourtListener URL or Legal Data Hunter document link).

- Distinguish **holding** from **dicta** from **procedural posture**; quote the operative

language rather than paraphrasing it.

- Flag confidence: **GREEN** (retrieved, read, supports the proposition), **YELLOW**

(retrieved but support is partial/arguable), **RED** (could not retrieve — do not cite).

Hand RED/YELLOW items to the state's `*-fact-check` skill before anything is filed.

- When research feeds a filing, repeat the NOT LEGAL ADVICE disclaimer in the deliverable.

## If the tools are missing

Both servers require a **one-time sign-in**: CourtListener uses OAuth with a free

CourtListener account; Legal Data Hunter uses GitHub/Google sign-in. If their tools are not

available in the session, tell the user to run `/mcp` in Claude Code and complete

authentication — do not silently fall back to memory. Hard fallbacks: CourtListener's REST

API (token-based) and the per-state `legal-data-apis.md` index of official sources.

## Composition

Composes with every state plugin's `*-fact-check` skill (this skill finds and verifies

authority; fact-check audits the finished draft), the `*-consumer-debt` and other

subject-matter bundles (which describe what to search for), the consumer FCRA skills in this

plugin, and `immigration-case-law` in `claude-legal-immigration-laws` (the

immigration-flavored counterpart).