A Claude skill that prepares a lawyer for an adversarial proceeding the way Neal Katyal's prep team prepared him for a SCOTUS argument.

Profile the tribunal from prior opinions and questions. Predict the specific questions you will face. Map narrow "escape routes" each judge can walk through without abandoning their commitments. Spar adversarially until only the strongest answers survive.

The model is the sparring partner. The human still wins the case.

In November 2025 Neal Katyal argued the SCOTUS tariffs case. His prep partner, "Harvey," was trained on every question every justice had asked in 25 years and every opinion they had written. Harvey predicted the bench near-verbatim and surfaced the narrow door the Chief Justice walked through.

The technique is reproducible. The skill encodes it.

You give Claude a hearing context: case, posture, the named tribunal members, the lawyer's draft brief or theory of the case. The skill runs five phases:

1. Profile the tribunal. One-page bullet profile per decision-maker — doctrinal commitments, recurring questions, tells in concurrences, institutional concerns, coalition behavior, phrasing tics.
2. Predict the bench. 3–7 questions per judge, each annotated with the doctrinal hook, the worry behind the question, and an attack-rank.
3. Map escape routes. For each plausibly hostile judge, find the narrow holding they could sign onto without abandoning a prior commitment. Hand them the door open.
4. Spar adversarially. Round-trip the lawyer's draft answers against the worst question on the bench. Iterate until the answer survives.
5. End with the human. Close every session by reminding the lawyer what only they can do at the podium: listen, connect, adjust tone, see the worry behind the question.

See examples/example-bench.md for the canonical output.

• Appellate oral argument prep
• Motion hearing prep
• Evidentiary hearing prep
• Deposition prep (taking or defending)
• Arbitration / mediation prep with named arbitrators
• Stress-testing a brief before filing

• The decision-maker is unknown or won't ask questions (paper-only motions).
• Pro se litigants in jurisdictions with hostile bench norms — the predicted-question framing can backfire.
• Anything where the prep work substitutes for the lawyer's actual judgment at the podium. The skill is a sparring partner, not a script.

git clone https://github.com/sboghossian/oral-argument ~/.claude/skills/oral-argument

That's it. Claude Code picks it up automatically. Trigger with /oral-argument or any of the trigger phrases (prep for argument, judge profile, predict questions, bench prep, …).

A generic LLM will give you confident, plausible predictions — and miss the narrow door.

This skill is built around five operating principles that prevent that failure mode:

1. Predictability is integrity, not weakness. Predictions are framed as respect for the judge's stated commitments, never as gotchas.
2. No parroting. The output is raw material the lawyer absorbs and re-delivers; recitation loses.
3. Find the door, don't push. The strongest move is to identify the narrowest ground a skeptical judge could rule your way while staying consistent with everything they've ever said.
4. Adversarial by default. Every argument is wrong until it survives the worst question on the bench.
5. End with the human. Always close by reminding the lawyer what only they can do live.

Read SKILL.md for the full methodology.

Pre-publish. v0.1.0 SKILL.md is stable; canonical example pending. Public-repo scaffolding complete; awaiting decision on hosted demo.

[TBD — contributor must add LICENSE before publishing publicly. Default stance is AGPL-3.0 to match Stephane's other public legal-AI skills.]

• forkcast — same author, world-model scaffolding for software project planning. The /lecun-world-model stance applied to code instead of court.
• awesome-legal-skills — community catalog of legal AI skills; this skill is a candidate for the catalog under oral-argument-stephane/.