Last Verified: October 13, 2026 — read the skill’s source (v1.0.0) and ran its plan command
Ask Claude whether your idea is good and it usually says yes. court is a free Claude Code skill that stops that: a prosecutor Claude argues your idea fails, a defense Claude argues it works, 12 juror Claudes vote on their own, and a judge Claude writes the verdict, plus the exact changes that would flip it.
Below: install it, check it’s in place, run your first trial, read the verdict, and seven cases you can paste today.
In this guide: Quick Start · Step 1 check the install · Step 2 your first trial · Step 3 read the verdict · Step 4 seven cases to put on trial · how the trial runs · fixes · FAQ
Quick Start
- What it is: a skill folder that adds a
/courtcommand to Claude Code. No app, no website. - Who made it: Alex Chen (@nocodealex), an independent developer. Not made by Anthropic.
- Cost: free and MIT-licensed. No signup, no API key. Each trial uses your normal Claude plan (17 sub-agent calls for a full jury).
- You need: Claude Code, Git and Python 3.
Paste into your terminal:
git clone https://github.com/alexyc9381/court-skill
cp -r court-skill/skills/court ~/.claude/skills/Rather let Claude do it?
Paste this into Claude Code instead and it installs and checks the skill for you:
https://github.com/alexyc9381/court-skill
Install this skill, then confirm /court works.Other ways to install
Skills CLI: npx skills add alexyc9381/court-skill. As a plugin: /plugin marketplace add alexyc9381/court-skill then /plugin install court-skill@court-skill (the command is then /court-skill:court).
On this page (8 sections)
Step 1. Check the Install
1a. Git and Python are there
~1 min
git --version
python3 --versionTwo version numbers means you’re set.
Windows: if python3 opens the Microsoft Store or isn’t found, install Python from python.org. During a trial you can also tell Claude use python instead of python3.
1b. The skill files are in place
~1 min
ls ~/.claude/skills/court
python3 ~/.claude/skills/court/court.py planYou should see the three files and the size of a full trial (this is our output from October 13):
jury of 12: 2 opening statements, 2 rebuttals, 12 jurors, 1 judge = 17 sub-agent calls
Then type / in Claude Code: /court should be in the list. If it isn’t, exit and start a new session.
court.py plan.Step 2. Run Your First Trial
2a. Open a folder for trials
~30 sec
The skill never edits your files; it writes everything into a .court/ folder wherever you run it. A dedicated folder keeps that tidy:
mkdir -p ~/court-trials
cd ~/court-trials
claude2b. Turn on accept-edits mode
~5 sec
Press Shift+Tab until the status line says accept edits on. Otherwise Claude asks you to approve every one of the ~17 files the jury writes.
2c. Write the case so a stranger could judge it
~2 min
The prosecutor, defense and jurors can’t see your conversation. They only get the case file, so the facts have to be in the case.
❌ /court is my hot sauce business a good idea?
✅ State it as a plan, then give the numbers:
/court Quit my job to sell homemade hot sauce online full time. I have $10,000 saved, sold 60 bottles at $12 at two farmers markets this summer, and my rent is $1,400 a month.It worked if Claude first tells you the size of the trial (for example 12 jurors, 17 sub-agent calls) and then starts the openings.
| Add after /court | Jurors | Sub-agent calls |
|---|---|---|
| (nothing) | 12 | 17 |
--quick | 6 | 11 |
--jury 3 | 3 | 8 |
A full trial is 17 sub-agent calls on your plan
Every juror is a separate sub-agent call that counts toward your Claude usage. Near your limit? Start with --quick (11 calls). --jury N takes any number from 3 to 12; --seed S fixes which jurors sit on a smaller jury so you can rerun the same panel.
Step 3. Read the Verdict
When the trial ends you get one file, .court/<run>/VERDICT.md. Read it in this order:
| Section | What it tells you | What to do with it |
|---|---|---|
| The verdict line | GUILTY or NOT GUILTY and the vote count | A close vote means the idea is fixable; a sweep means rethink it |
| How the jury voted | Every juror’s vote and one-line reason | Look for the same objection from different jurors |
| Charges that stuck | The arguments that won, with how many jurors agreed | These are your real risks, ranked |
| What the defense got right | The strongest case for your idea | Keep these; they’re what’s already working |
| Conditions for acquittal | Specific changes that would flip votes | Your to-do list before you try again |
| The sentence | The judge’s plain-language order | The one thing to do next |
Here’s part of the repo’s own example, a 3-juror trial of “Quit my job as a nurse to sell handmade candles on Etsy full time”:
Charges that stuck:
1. The plan gives up the paycheck first and risks savings worth only five months of rent (3 jurors).
2. About 13 sales a month is far short of the roughly 57 needed to cover rent (2 jurors).
The sentence: do not hand in your notice yet.
Excerpt from examples/candles/VERDICT.md, court skill by Alex Chen (@nocodealex), MIT.
- Met the conditions? Update the facts in the case and run a new trial.
- Disagree with the verdict? Don’t ask Claude to overturn it; the skill is built not to. Ask for a retrial with the conditions written into the new case.
A second opinion, not advice
The jury is 12 language models reading only your case file. Use the verdict as a structured second opinion, not legal, financial or professional advice.
Step 4. Seven Cases to Put on Trial
Write a claim, not a question
Jurors vote guilty or not guilty on a statement. “Is X good?” gives them nothing to vote on; “Doing X will work, here are the numbers” does.
1. Quitting for a side business
/court Quit my job to run my [side business] full time. Facts: [monthly revenue for the last 3 months], [savings], [monthly costs], [what I'd do if it fails].2. Raising your price
/court Raise my price from [$X] to [$Y] for [product]. Facts: [current customers], [how many bought last month], [what competitors charge].3. Launching a product
/court Launch [product] to [audience] on [date]. Facts: [who asked for it], [pre-orders or waitlist size], [budget], [what's not built yet].4. Making a first hire
/court Hire a [role] at [salary] next month. Facts: [revenue], [hours I spend on this work each week], [months of runway after the hire].5. Changing your content
/court Switch my [channel/newsletter] from [current topic] to [new topic]. Facts: [current audience size], [best and worst results], [why I want to switch].6. A big personal move
/court Move from [city] to [city] this year. Facts: [job situation], [cost of living change], [who I know there], [what I'd give up].7. A claim in your pitch or post
/court This claim is true and will hold up with customers: "[your claim]". Evidence: [numbers, sources, tests you ran].How the Trial Runs
| Phase | Who | Calls (full jury) |
|---|---|---|
| Opening statements | Prosecutor and defense each build their strongest honest case | 2 |
| Rebuttals | Each side answers the other’s points | 2 |
| Jury | 12 jurors read the full record and vote alone; no juror sees another’s vote | 12 |
| Judge | Reads the votes, writes what stuck, what the defense got right, the conditions and the sentence | 1 |
A small Python script (court.py) does the bookkeeping: it writes each agent’s brief, checks every output exists and counts the votes. Claude only runs the trial. It never argues a side, votes, or edits the verdict.
The 12 jurors each judge through a different lens:
The Accountant · The First-Time Customer · The Competitor · The Retired Judge · The Teenager · The Engineer · The Small-Business Owner · The Statistician · The Parent · The Investor · The Designer · The Lawyer
With a smaller jury (--quick or --jury N), a random subset sits, recorded so --seed can repeat it.
If Something Goes Wrong
| You see | Why | Fix |
|---|---|---|
/court isn’t in the command list | The session started before the install, or the folder is in the wrong place | Check Step 1b, then exit Claude Code and start a new session |
python3: command not found | Python isn’t installed, or Windows only has python | Install Python 3, or tell Claude to use python instead |
cp: ... No such file or directory | ~/.claude/skills doesn’t exist yet | Run mkdir -p ~/.claude/skills, then the copy again |
| An approval prompt for every file | The default permission mode asks before each write | Shift+Tab to accept-edits mode before you start (Step 2b) |
| Jurors argue about facts you never gave | Sub-agents only see the case file, not your chat | Put the numbers in the case and rerun |
A juror shows NO VOTE | That juror failed twice | Nothing to fix, the trial still finishes. Rerun if the vote is close |
| You hit your usage limit mid-trial | 17 sub-agent calls in one go | Use --quick or --jury 3 next time |
To remove it: rm -rf ~/.claude/skills/court
Official Sources
- alexyc9381/court-skill on GitHub (README, source, the candles example)
- The author’s own step-by-step guide at chen.media
FAQ
Is it really free?
The skill is free and MIT-licensed with no signup or API key. The trial itself runs on your Claude Code plan: a full jury is 17 sub-agent calls, --quick is 11.
Does it touch my project files?
No. Everything it writes goes into .court/ in the folder you run it from.
Can it be wrong?
Yes. The jurors are language models reading only what you wrote. A thin case gets a shallow verdict; the value is in the specific objections and conditions, not the vote alone.
Why not just ask Claude to be harsh?
One Claude playing critic still sees your framing and its own earlier answer. Here, separate agents argue each side, and jurors vote without seeing each other, so nobody can just go along with the room.
Related tools
More free add-ons in our 160+ subagents guide and 5 Claude Code settings to change first.
/court skill by Alex Chen (@nocodealex): github.com/alexyc9381/court-skill