How to Stop Claude AI Inventing Cases You're About to Cite (5 Cues)

#ClaudeForLawyers #LegalAI #AIHallucinations A New York federal court, 2023. Two practicing lawyers filed a brief containing six cases that did not exist, were sanctioned five thousand dollars, and watched the story go around the world within forty-eight hours. Mata v. Avianca became the most famous citation that never was. Three years later, the pattern has not stopped. It has spread. Fabrication is the single most career-defining risk in legal A·I use today. The panel has developed five prompt-level cues that materially reduce the rate at which Claude invents authority, before any verification work happens at all. The cues do not eliminate the need for verification. Nothing eliminates the need for verification. But they shift the failure rate before the verification step begins, which means fewer fabricated citations in the draft and lower total risk on every brief filed. In this episode the panel walks through each cue in detail: source restriction, jurisdiction anchoring, citation discipline, negative instruction, and verification handoff. Plus the combined prompt template that runs all five together with about ninety words of overhead per research request. Plus three honest limits on what the cues do not do, because overstating what the protocol accomplishes is itself a risk to the practitioner who trusts it too completely. This is one of the durable Foundations skills. It does not depend on any specific plugin, connector, or product surface. The protocol applies whether you're running Claude through the consumer app, the API, or inside Claude for Legal. For practitioners using the Claude for Legal release, the research connector layer makes the first cue substantially stronger, because the model is reading from Westlaw and CourtListener directly. The other four cues apply equally inside and outside the Claude for Legal environment. The full five-cue research prompt template is in the pinned comment. Copy it. Save it in your working drafts. Paste it at the top of every research request you give Claude on a live matter. — Resources referenced in this episode: · Channel — Foundations No. 06 (How to Verify Every Citation with Claude AI) — the downstream verification protocol the cues compound with · Channel — Unpacked No. 04 (The Research Connector Workflow) — Westlaw, Practical Law, and CourtListener inside Claude, which make the source-restriction cue substantially stronger · Mata v. Avianca, 22-cv-1461 (S.D.N.Y. 2023) — the reference case for legal A·I fabrication sanctions — Coming next in Foundations: No. 08 — How to Quickly Prep a Witness with Claude AI (90 Minutes til Showtime) No. 09 — How to Brief a Partner on a Matter You've Never Seen with Claude AI — For partnership, sponsorship, and creator collaboration enquiries: [email protected] For bespoke Claude training tailored to your firm or in-house team: [email protected] — DISCLAIMER. This channel is for educational purposes only and does not constitute legal advice. The panel framing represents the collective judgement of a panel of senior practitioners with decades of combined experience across multiple practice areas. Claude is an AI assistant produced by Anthropic; this channel is editorially independent and is not a vendor channel. All workflows discussed should be reviewed against your jurisdiction's professional conduct rules and your firm's information governance policies before deployment on a live matter.