Court

Court is a skill for people who are tired of agreeable answers. Instead of one polite reply, the idea goes on trial. A prosecutor builds the case against it, a defense builds the case for it, and a jury of independent jurors who never see each other votes. A judge then writes the closing sections. The agent running the skill acts as the clerk of the court: it never argues, never votes and never softens the verdict.

All bookkeeping goes through a script, scripts/court.py in the skill's folder. State lives in .court//state.json in the current directory, and each command after init finds it through .court/LATEST. The trial runs in phases: opening, rebuttal, jury, judge. The command "next" reads the files on disk and reports the next phase and the exact command to run. Finally "render" counts the votes and writes VERDICT.md into the run folder.

Use it when the user wants pushback instead of agreement, says the assistant agrees with everything, asks "is this a good idea?" about a plan, a purchase, a post, a code plan or a claim, or types "court", "put it on trial" or "/court". If the skill fires on its own without a request for a trial, it asks once before starting and offers the quick option.

Do not rely on it for legal, financial or professional advice; the source calls the verdict a structured second opinion. Vague cases get vague verdicts, because the jury only knows what the user wrote. A full trial is 17 sub-agent calls (11 with the quick option) and each counts toward the user's plan usage, so it is not a lightweight check. It runs in an environment that can spawn sub-agents and call its script through the skill directory variable.

The input is the case, typed after /court, plus optional flags: --jury N sets 3 to 12 jurors (default 12), --quick uses 6 jurors, and --seed S fixes which jurors sit on a smaller jury (default random, and recorded). If the argument is blank or still a placeholder, the case is taken from the idea or plan the user most recently wanted an opinion on. The tool is invoked like this:

python3 "${CLAUDE_SKILL_DIR}/scripts/court.py" <command>

The clerk first runs plan and tells the user the size of the trial in one line. It then writes .court/case.md so it stands on its own, because sub-agents cannot see the conversation. The file starts with the case in one sentence, in the user's words where possible, followed by every relevant fact the user gave: numbers, audience, budget, timeline, what was tried, constraints. The clerk adds no facts and no opinion, since a leaning case file decides the trial before it starts.

Each phase writes briefs, then launches one wave of sub-agents at a time. If a juror fails twice, the line NO VOTE is written to its file and the trial moves on; the clerk never writes a vote or statement itself. The output is the verdict line, a vote table and the judge's sections, shown exactly as written, plus at most two lines from the clerk: the first condition for acquittal and the path to the full trial.

The skill works with sub-agents spawned from the host and writes into a .court folder in the current directory. In default permission mode this means one approval per file, so the skill suggests accept-edits mode for the trial, but it never changes the user's settings itself.

Founders, makers and anyone about to commit money, time or reputation to a plan: quitting a job to sell candles, buying something big, publishing a post, or approving a code plan. If the user disagrees with the verdict, the clerk offers a retrial with the conditions for acquittal written into a new case file, and does not argue for either side. Example: "/court --quick The plan you just gave me is the right one."