Make 100 Versions of Claude Fight for Your Best Answer

Here's the problem with asking Claude for one answer.

You ask. It gives you something okay. You rephrase. It gives you something a little better. You beg. It gives you something close. After three or four rounds you settle, because you're tired, and "close" feels good enough.

The Arena Skill fixes that. You say one word and Claude launches a hundred versions of itself, each taking a different approach to your task, all fighting head to head until one champion is left. Then it builds a remix out of the best ideas from the losers and makes that remix fight the champion. If the remix wins, that's what you get.

It works for anything where "which version is best?" has an answer a judge can reason about. Headlines. Ads. Emails. Offers. Names. Scripts. Plans.

What you need

The Claude desktop app or Claude Code. The skill has to be saved as a skill (a SKILL.md file), which you can do from either one. A regular chat in your browser won't let you install it the same way.

A paid plan runs the full tournament faster and without hitting limits, because it spawns many sub-agents. The free version works for small runs.

About ten minutes to install it the first time. After that, using it takes one word.

Step 1: Install the skill

Open Claude Code or the Claude desktop app and paste this whole prompt in. It tells Claude to save the skill exactly as written, word for word, without rewriting, shortening, or "improving" anything:

CopyAdd a skill to my Claude account. The full text is below. Save it exactly as written, word for word. Do not rewrite, shorten, summarize, or improve anything.

Name: arena
File: SKILL.md
Content: everything between the line START ARENA SKILL and the line END ARENA SKILL.

How to do it:
1. Use your skill-creator or skill-saving tool to create the skill and show me the save card.
2. The START and END lines are markers for you. Do not put them inside the skill file.
3. If you have no tool that can save skills, reply with the skill text alone in one code block, and tell me in one sentence that I can paste it into my skills settings or into a project's instructions.
4. When it is saved, tell me the skill name and one short line on how to start it. Do not run it yet.

After it is saved, I will type "arena" followed by my task. For a first run I will use a small field, around 8 fighters.

START ARENA SKILL
---
name: arena
description: Runs a tournament where many AI agents each write their own version of the same deliverable (a headline, ad, email, offer, name, script, or plan), then judges knock versions out head to head until one champion is left, followed by a remix of the best ideas that has to beat the champion to win. Use this whenever the user says "arena", asks agents to compete, fight, battle, or go head to head, or wants the strongest possible version of something out of a large field of options, even if they never say "tournament".
---

ARENA

The arena turns one task into a contest between many agents. One attempt gives you one idea. A hundred attempts, judged against each other with a fixed rubric, give you the strongest idea plus the best pieces of the losers. It works because competition forces variety at the start and honest comparison at the end.

"Fight" means a contest of outputs. Each agent writes its best entry and judges decide who beats whom. The arena fits any task where "which version is best?" has an answer a judge can reason about. If the task itself is harmful or deceptive, decline it the same way you would outside the arena.

STEP 0: CHECK WHAT YOU CAN ACTUALLY DO

Before anything else, check your tools.

Full mode: you have a tool that launches sub-agents (often called Agent or Task) and you can write files. Run the full tournament below.

Lite mode: you have no sub-agent tool. Run a small arena inside this chat instead. Write 8 to 12 entries yourself, one per corner, each in its own block. Number them, strip the corner labels, then judge them with the same rubric and the same bracket rules. Tell the user in one plain sentence that this is lite mode and that the entries come from one mind, so variety is lower than a full run.

If you have sub-agents but cannot write files, run full mode and keep the entries in the conversation instead of in files.

Never say more agents ran than actually ran. The final report states the real count.

STEP 1: BEFORE THE BELL

Settle these six things. Ask only for what is missing, in one short message, and offer a default for each so the user can just say "go".

1. The task. What are the fighters producing? Example: "a 10-word headline for a service that follows up with law firm leads within 60 seconds."
2. The reader and the goal. Who reads the winner, and what should they do after reading it? Fighters write badly without this.
3. The rubric. Three to five criteria, weighted to add up to 100, most important first. If the user gives none, propose one in a single line and proceed. Judges with no rubric drift toward whatever sounds smoothest. Starter rubrics:
   Headline or ad: stops the scroll 30, clear promise 25, believable 25, fits the reader 20.
   Email or DM: gets opened 25, gets a reply 35, sounds like a person 25, short enough 15.
   Offer: value is obvious 30, risk is reversed 25, specific 25, easy to say yes to 20.
4. The field size. Default 100 in full mode. If the user names a number, use it. For a first run on any new task, suggest 8 fighters so a weak rubric gets caught cheaply.
5. The entry limit. Cap every entry, for example 120 words. Short entries keep large fields affordable and stop judges from rewarding length.
6. The hard rules. Anything entries must include or avoid: required phrases, banned claims, tone, legal limits.

Confidentiality: send fighters only what they need to do the task. Do not pass client names, private numbers, or personal details unless the user says to.

Cost and the go-ahead: before launching, state the plan in two lines: field size, stages, and the rough number of agent calls (see COST below). If the user asked for a specific field size, treat that as the go-ahead and launch. If they did not, and the field is above 24, ask once for a yes before launching.

STEP 2: BUILD THE FIELD

100 copies of the same prompt produce 100 near-identical entries. Prevent that by giving each fighter its own corner: a strategy, angle, or persona that pushes it in a different direction.

Write 12 to 20 corners that fit this task. Generic starters to adapt: blunt, story-first, numbers-first, contrarian, skeptical-buyer view, urgency, humor, plain-spoken neighbor, authority, minimalist, curiosity gap, emotional. Include at least two odd corners the user would not have thought of. Assign corners round-robin so each one gets several shots.

Also write a short banned-cliche list for the task (for example "game-changer", "unlock", "in today's fast-paced world") and give it to every fighter. This keeps the field from collapsing into the same stock phrases.

Number fighters 001 to 100. Judges see numbers only, never corners.

Each fighter's prompt contains: the task, the reader and goal, the rubric, the entry limit, the hard rules, its corner, the banned list, and this instruction: commit fully to your corner, write only the entry, and add nothing else. Fighters do not know about the other fighters.

Truth rule for every entry: no invented statistics, testimonials, awards, case results, credentials, or client names. Use only facts the user supplied. In regulated fields such as law, medicine, finance, and real estate, entries must not promise outcomes or guarantee results, and the final report should tell the user to have a qualified professional review the winner before it is published. Give that as a flag, not as legal advice.

STAGE 1: THE BRAWL

Launch fighters in parallel batches. If the environment rejects a large batch, use smaller ones and keep going. Use a fast, inexpensive model for fighters if you can choose one. Volume counts for more than polish here.

Each fighter writes its entry to arena-workspace/round-0/entry-NNN.md, containing only the entry.

When the batches finish, count the files. Re-run any missing fighter once. If one fails twice, note it in the report and carry on. Do not read all the entries yourself. The judges read them. This keeps your own context clear.

Disqualification: a judge disqualifies any entry that runs over the limit, breaks a hard rule, or contains an invented claim, and writes the reason down. Disqualified entries do not advance.

STAGE 2: PODS

If the field is 16 or smaller, skip pods and go straight to the knockout with a random pairing.

Otherwise, split the field into pods of about 10. Shuffle entries into pods at random. If a shell is available, use a script to shuffle. If not, deal them out like cards so neighbors land in different pods. Pods built from neighboring numbers would carry the corner pattern into the judging.

One judge agent per pod. Use a stronger model for judges if you can choose one, because judging is where quality is decided. Each judge:
1. Scores every entry on every rubric criterion from 1 to 10, and computes the weighted total.
2. Picks the top 2 by weighted total. Ties go to the entry with the higher score on the most important criterion.
3. Writes arena-workspace/round-1/pod-NN.md with the full score table, the two winners, one line for each citing a rubric criterion, any disqualifications, and one strong idea from an eliminated entry worth stealing.

STAGE 3: KNOCKOUT

Pod winners enter a single-elimination bracket, paired at random. For 100 fighters the path is 20, 10, 5, 3, 2, 1. With an odd number of survivors, the entry with the highest pod score gets a bye.

Each match uses one judge. Show the two entries as A and B with the order randomized. The judge returns the winner, a one-line reason that quotes a rubric criterion, the margin (narrow, clear, or lopsided), and any idea from the loser worth keeping.

Semifinals and the final get extra protection against position bias. Run two judges, with the A and B order swapped for the second. If they disagree, a third judge decides.

Duel mode (optional, offer it when the user wants a real fight): in the semifinals and final, each fighter first reads its opponent's entry and writes a short attack on its weakest point and a short defense. The judge sees both entries and both attacks. This adds about 2 calls per match, so use it only in the last rounds.

STAGE 4: REMIX AND REMATCH

Write one remix that combines the champion with the best stolen ideas, inside the entry limit and under the same truth rule. Then make it fight: run the champion against the remix as one more blind match with a fresh judge.

If the remix wins, it becomes the recommended version. If it loses, the champion stays the recommendation and the remix is listed as an alternative. Report the result honestly either way. A remix that never fought is only an opinion.

STAGE 5: THE REPORT

Write arena-workspace/report.txt as plain text with no decoration. Include:
1. The recommended entry, in full, and whether it is the original champion or the remix.
2. Why it won, in two or three sentences drawn from the judges' reasons.
3. The runner-up and the other finalists, one line each.
4. The best ideas from the fallen, collected from the judges and deduplicated.
5. A fairness check: did the winners cluster in one corner, did the semifinal and final judges disagree, and what that suggests about the rubric. If every pod winner shares one corner, say so. It usually means the rubric favors one style.
6. The real numbers: mode used, field size, stages, disqualifications, failed fighters, and agent calls actually used.
7. If the field is regulated, the professional-review flag.

Send the report to the user as a file if you can. In chat, give the recommended entry, whether the remix beat the champion, and one line on what to try next. Do not recap every round. Write the chat reply in plain language for someone who has never heard of a tournament bracket.

If you cannot write files, put the same report in the chat, kept short.

COST

Approximate agent calls for a full run: fighters, plus one judge per pod, plus one judge per knockout match, plus 3 extra for the double-judged semifinals and final, plus 1 for the remix match.
100 fighters: 100 + 10 + 19 + 3 + 1, about 133 calls.
24 fighters: 24 + 2 + 3 + 3 + 1, about 33 calls (one pod round is skipped when the field is small, so treat this as a ceiling).
8 fighters: 8 + 7 + 3 + 1, about 19 calls.
Duel mode adds about 2 calls per duel match. Say the figure in the pre-launch plan so the cost never surprises the user.

FAIRNESS RULES

Judges see numbers only, never corners or authors. Match order is randomized. Judges quote a rubric criterion for every decision, and a reason with no criterion behind it gets rewritten. Judges never see each other's verdicts before deciding. Every file stays in arena-workspace so the user can audit any match. If something went wrong in a stage, say so in the report instead of hiding it.

AFTER THE RUN

Offer one of two next moves, in one line: a rematch with a tightened rubric based on what the fairness check showed, or a polish round where a small field of 8 fighters each tries to improve the winner on its weakest rubric criterion.
END ARENA SKILL

Claude will walk you through the save. When it's done, it tells you the skill name and one line on how to start it, then stops. Don't run it yet.

Important: Don't edit the skill text. The whole thing is written to be saved exactly as-is. If Claude "improves" it, the tournament stops working the way it's supposed to.

If Claude says it has no tool that can save skills, it will hand you the skill text in one block instead. Paste that into your skills settings or into a project's instructions.

Step 2: Do a small test run first

Before you throw a hundred agents at something, run a quick eight-fighter test. It catches a bad rubric cheaply.

Type this in a new chat:

Copyarena — write a 10-word headline for [your offer]. 8 fighters.

Claude asks you for anything it's missing — usually the rubric (the criteria the judges score against), the reader and goal, and any hard rules. It offers a default for each so you can just say "go." Give it what it needs, and it runs.

Eight agents each take a different angle, judges pick the best, and you get a champion plus a remix that had to beat the champion to win.

Step 3: Read the report

You'll get a report with seven things in it:

• The recommended entry, in full, and whether it's the original champion or the remix
• Why it won — two or three sentences from the judges
• The runner-up and the other finalists — one line each
• The best ideas from the fallen — the strong pieces the judges stole from eliminated entries
• A fairness check — whether the winners clustered in one corner, whether the final judges disagreed, and what that suggests about your rubric
• The real numbers — the mode it ran in, the field size, how many fighters failed or were disqualified, and the agent calls actually used
• A professional-review flag if your field is regulated (law, medicine, finance, real estate), reminding you to have a qualified professional check the winner before you publish it

The remix is often better than the champion. Read it last. And read the fairness check — if every winner came from one corner, your rubric probably favors one style, and tightening it will change the result.

Step 4: Run it for real

Once the test run looks right, go full size:

Copyarena — [your task]. 100 fighters.

For 100 fighters it's about 133 agent calls. Claude tells you the plan before it launches, so the cost never sneaks up on you. If that feels like a lot, pick a smaller number. Twenty-four works (about 33 calls). Eight works (about 19). The skill scales to whatever you name.

Lite mode

If you're in a chat with no sub-agent tool, the skill doesn't give up. It runs a small arena right there in the conversation — it writes 8 to 12 entries itself, one per angle, strips the labels, and judges them with the same rubric and bracket rules. It tells you plainly that this is lite mode and that the entries come from one mind, so there's less variety than a full run. It's a good way to try the idea before you install it somewhere that can run the full thing.

Where this pays off

The arena is built for tasks where one answer isn't enough and you want the best out of a wide field:

• Headlines and subject lines
• Ads and sales offers
• Emails and DMs
• Names for a product, a campaign, a business
• Scripts for calls and videos
• Strategies and plans where you want options compared honestly

It's not for everything. If you need one correct factual answer, just ask. The arena is for the judgment calls — the things you'd otherwise guess at, or settle on after three rounds of rephrasing.

What to do next

Install the skill today. Run the eight-fighter test on a real headline or offer you're working on right now.

Then the next time you're about to rephrase a request for the fourth time, stop. Type "arena" instead and let a hundred versions fight it out. You'll never go back to begging.