From Research Question to Cite-Checked Brief Section: The CourtListener + Midpage Kit

A drafting and review aid for lawyers and compliance teams. It does not give legal advice, and a qualified lawyer should check anything it produces before it is relied on or sent.

A brief section usually starts with a research question and a deadline. The slow part is the loop between finding the cases, reading them, writing the argument, and then going back to confirm every cite says what the draft says it does. AI is now good at the finding and the first draft. It is still risky at the citations, so this kit puts a check between the draft and the attorney.

This kit connects your assistant to CourtListener, Free Law Project's free case-law database, through its official MCP connector. It adds Midpage where a paid subscription earns its keep, drafts one section with Anthropic's open-source Brief Section Drafter skill, and runs every citation through a free checker before a lawyer reviews it. You end with a draft section, a list of every authority it relies on, and a citation report.

What you need: Claude with the litigation plugin from Claude for Legal installed (it runs in Claude Cowork or Claude Code), and a free CourtListener account. Midpage is optional and paid.

Important: this is a workflow guide, not legal advice. The tools here help you find, organize, and draft. They don't decide what the law is, and they can be wrong. A licensed attorney must read every authority, verify every citation and quote, and take responsibility for anything that gets filed. Follow your firm's rules and any court orders on AI use, and on which tools may see client material.

The steps, estimated times, and directory pages are in the kit card above. Every time in this kit is an estimate. Yours will depend on the issue, the record, and how much of the research you already have.

Which tools work where

Each publisher's own setup page says which assistants are supported, as of October 22, 2026:

The research step works in ChatGPT or Cursor too. The drafting skill is Claude-only, so this kit runs in Claude end to end.

Before you start: set up once (estimated 20 minutes)

1. Install the litigation plugin from the Claude for Legal repository and run its setup interview. The drafting skill reads your case theory and house style (citation format, structure, tone, length) from the configuration that interview saves.
2. Connect CourtListener. The litigation plugin already lists the CourtListener connector, so you authorize it with your CourtListener account. In plain Claude, add it from the connector directory. In Claude Code, Free Law Project's command is claude mcp add --transport http courtlistener https://mcp.courtlistener.com/. Any other MCP client takes https://mcp.courtlistener.com/ as a custom connector URL with OAuth.
3. Optional: add Discernment Nudge, a general-purpose Anthropic skill used in Step 5.

Step 1: Frame the issue (estimated 20 minutes)

The research is only as good as the question. Before you search, write down three things: the precise question the section has to answer, the procedural posture and the standard the court applies, and the jurisdiction whose authority binds you.

If the matter is new to the plugin, start with Matter Intake. The Brief Section Drafter checks the plugin's matter log first and refuses to draft on a matter that hasn't been taken in, because intake is where the conflicts check runs.

Ask:

New matter: intake [matter name]. We represent [party]. Here's the complaint and the scheduling order.

Then state the question in one sentence, in the chat, before any research. For example: "Does [rule] require [element] at the pleading stage in the [circuit]?"

Step 2: Research the case law (estimated 60 minutes)

With the CourtListener MCP connected, the assistant searches CourtListener's live data instead of relying on what the model remembers. Per Free Law Project's docs, that covers federal and state opinions, PACER dockets, the citation network (what a case cites and what cites it), oral argument audio, judge data, keyword and semantic search, alerts, and citation verification.

Ask:

Using CourtListener, find [jurisdiction] appellate decisions from the last ten years on [question from Step 1]. For each case, give the citation, the holding in one sentence, and a pinpoint to the passage that supports it. Then list the cases that cite the two strongest results.

Read the cases it returns, not just the summaries. A search result is a lead, not an authority.

Where Midpage adds something. Midpage is a paid alternative or second source. Its docs describe an AI citator that labels later treatment Negative, Caution, or Neutral with a supporting quotation, a tool that reads one opinion against your question and returns supporting quotations, and statutes, regulations, and agency guidance alongside the case law. Midpage's own docs say the treatment list "is not a complete good-law check" and that neutral treatment is not proof a case is still good law, so it doesn't replace your citator.

Midpage is a subscription. Its pricing page lists the Professional plan at $100 per seat per month, billed monthly, after a two-week free trial, and its custom-connector page says the connector is "for individual attorney use only." Check that your plan and your firm's tool policy cover your use before you connect it.

Step 3: Draft the section (estimated 45 minutes)

Brief Section Drafter writes one section at a time: statement of facts, standard of review, argument, or conclusion. It first asks whether the section is for a written submission or oral argument. It drafts in the house style from your setup and checks that the section fits your case theory.

Ask:

Draft argument section II on [question from Step 1] for [matter], using the CourtListener research above. Written submission.

What to expect in the draft, per the skill file:

• Markers. Anything not confirmed is marked [VERIFY] (a fact), [UNCERTAIN] (a legal proposition), or [CITE NEEDED]. A draft with open markers is not final.
• Source tags. Every citation says where it came from: [CourtListener] or another research tool, [web search — verify], [model knowledge — verify], or [user provided]. Check the "verify" tags first.
• No silent gap-filling. If the research returned little for a point, the skill is written to say so and ask how you want to proceed instead of filling the gap from memory.
• Weak points flagged. When the authority cuts the other way, the skill is written to flag the argument as weak rather than dress it up.
• Drafting notes for the reviewer, including the authorities relied on and the open questions for the partner.

Your job in this step: don't delete the markers or the source tags. They are the checklist for Steps 4 and 5.

Step 4: Check every citation (estimated 20 minutes)

Run the draft through a citation checker before anyone treats it as close to final.

cite.review is the no-install option. It's a free, open-source web app, and you need a free CourtListener API token. Paste the section or upload it as DOCX, PDF, TXT, Markdown, or HTML. It checks cases against CourtListener, and federal statutes, regulations, and public laws against Cornell LII and GovInfo, then marks each citation Verified, Warning (found, but some fields don't match), or Not Found. An optional second step uses your own AI key to ask whether each verified case supports the proposition it's cited for.

Know the limits its docs state. Not Found doesn't always mean fabricated, because CourtListener has coverage gaps. State statutes need a manual check. For statutes it confirms the provision exists, not that it says what you claim. Per its docs, the document is parsed in your browser and Step 1 sends only citation strings. The optional support check sends excerpts to the AI provider you choose, so clear that with your firm's policy first.

Other options:

• Inside Claude: Free Law Project lists "verify every citation in this brief and flag any unknown citations" as a sample prompt for the CourtListener MCP.
• Self-hosted: the CourtListener Citation Validation MCP runs locally (Docker, or STDIO on Windows or Linux) for Claude Desktop and Claude Code. It inventories every citation and then checks each case citation against CourtListener. Its README notes that a "valid" result only confirms the citation is in CourtListener's index, not that it supports your proposition.
• For builders: eyecite is the open-source Python library CourtListener uses to find citations in text, including full and short case cites, statutes, law journals, supra, and id. The validation MCP above uses it.

Your job in this step: resolve every Warning and Not Found by pulling the source yourself.

Step 5: Attorney review (estimated 45 minutes)

No tool finishes this step. With Discernment Nudge installed, Claude can end a substantive draft with two or three pointed questions about what it just produced. The skill does this at most once per conversation and skips it when you've already asked for a review, so treat those questions as a starting point.

Then the reviewing attorney:

1. Reads every authority the section relies on, in full, and confirms each pinpoint supports the whole proposition it's attached to.
2. Confirms each case is still good law in the citator the firm relies on.
3. Checks every quotation word for word against the source.
4. Resolves or deletes every [VERIFY], [UNCERTAIN], and [CITE NEEDED] marker.
5. Checks length, format, and citation rules against the forum's local rules and the judge's standing orders, including any order on AI use.
6. Decides what gets filed. The skill doesn't file anything.

For background on the professional duties involved, see ABA Formal Opinion 512, the ABA's ethics guidance on lawyers using generative AI.

Want the formatting as files you can reuse? The Legal & Litigation Workflow Pack includes Cite-Check Prep, which pulls every citation into a verification table, and an AI-Use Disclosure Checklist for court AI orders. It formats the work. It doesn't verify citations or decide the law, and it doesn't include these free tools.

Common pitfalls

• Researching before you frame. A vague question gets a pile of loosely related cases.
• Citing from a summary. Summaries and search snippets aren't the opinion. Read the passage before you cite it.
• Treating "Verified" as "supports." A checker confirms that a citation exists. Only reading the case tells you whether it supports your point.
• Stripping the markers. Removing [VERIFY] or [model knowledge — verify] tags hides exactly what the reviewer needs to see.
• Treating a treatment signal as a full good-law check. Midpage's own docs say its treatment list isn't one.

Your research-to-brief checklist

• Install the litigation plugin, run setup, and connect CourtListener
• Intake the matter and write the question in one sentence
• Research in CourtListener, and Midpage if you subscribe, and read the cases
• Draft one section and keep every marker and source tag
• Run every citation through a checker and resolve each Warning and Not Found
• Attorney reads every authority, confirms good law, and decides what gets filed

Find these tools and more free legal tools in the YourAIKit directory.

YourAIKit products are workflow tools. They are not legal, tax, or financial advice.