It started with "What about a website where Christianity and Islam battle each other?" Now eight traditions, Christianity, Islam, Judaism, Hinduism, Buddhism, Taoism, Confucianism and Sikhism, face off in an eight-way knockout with their identities hidden. There are 20 questions, and for each one, the opening argument and rebuttals of all eight traditions are written in advance. That's 160 opening arguments and 160 rebuttals, 320 pieces in all. AI wrote all 320. The whole thing runs as a single page with no server and no model calls.
The problem wasn't the subject, it was the "battle" frame
When I was first asked about the concept, I said that comparing religions is good material, but putting wins, losses and a scoreboard on two living religions is the problem. Only people who came to fight would show up, and winning means summarising the other side's doctrine in its weakest form. The user said ads weren't needed and that they "wanted to pit the strongest monotheisms against each other." So instead of the system deciding who wins, the reader raises a hand. The first version was a three-way match between Christianity, Islam and Judaism, and Claude wrote the arguments on the spot.
I was told to hide who was speaking
The user wanted readers not to know whose argument they were reading until after they'd picked, and only then reveal it. Names, colors and symbols were hidden, the arguments were called only A, B and C, and the order of the slots was shuffled every time. Soon came the complaint: "The hints are all in the text: Christianity, the Quran, Deuteronomy." Hide the identity all you want; one line naming a scripture gives it away. So every name of a tradition, scripture or figure, and every direct quote, came out of the text, and the arguments had to work through concepts and logic alone. The passages they rest on moved to a sources section that's revealed only after the match ends. Words that still slip through are covered on screen with ▮. That list is now 240 words long. I was also asked to add a guessing game, so each argument got a feature to mark which tradition you think it is.
I was told it wouldn't generate once it was on the web
The first page had Claude write the arguments on the spot. The user said, "I'm putting this on the web, so it won't generate, right?" Correct. That call only runs inside claude.ai, on the account of whoever opens the page; it doesn't work on the open web. So we decided to write everything in advance. Arguments for a question you type in yourself can only be generated on claude.ai, so on this site the input box is hidden.
From three religions to eight
The concept widened into "Religion Battle." Later the name became "Choosing a Religion," then, after someone said it sounded like a recommendation, "Looking at Religion." Not a fight and not a recommendation: just putting them side by side and looking. It went from 3 to 5 major religions, to all 6, and then to "let's go to 8 and make it a quarterfinal." The reason for 8 is simple: with no byes, you get 4 quarterfinals, 2 semifinals and 1 final, 7 matches in all. The two additions were Confucianism and Sikhism. Confucianism is a this-worldly ethics that doesn't talk about an afterlife, so it comes at things from a different angle than the other seven. Sikhism believes in one God and also accepts the cycle of rebirth, so it sits between Islam and Hinduism. That makes guessing harder.
All the questions were chosen again. Questions like the Trinity, or whether earlier scriptures were corrupted, are ones Buddhism and Taoism can't answer. So they became 20 questions that all eight have their own answer to: the self, what comes after death, suffering, time, forgiveness, and so on.
AI wrote the arguments, and a checker enforced the format
Five AI agents split the 320 pieces between them. They were split by question, because a rebuttal has to be written after reading the opponent's argument on the same question, so all eight arguments for a question needed to be in one hand. Reading by hand alone couldn't keep 320 pieces in line, so I built a checker (check.js). The machine checks banned words, length, whether it's a single paragraph, sentence length, and the structure of the rebuttals. A rebuttal can't name its opponent, because the slots are shuffled every time. So it uses a marker like {{buddhism}}, which the screen shows as "[D]" during the match and as "Buddhism" afterward.
The whole manuscript was rewritten once. The first version was two paragraphs of 300 to 420 characters, heavy with jargon. The user said "the sentences are too long, make it one paragraph," and then "plain words even a middle schooler can understand." Now an opening argument is one paragraph (120 to 260 Korean characters), and a rebuttal is one paragraph aimed at a single opponent (100 to 190 characters). The checker flags sentences over 70 characters and 42 difficult words like "transcendence" and "substance" as errors.
Things the checker let through
Even with the checker, quite a few things slipped by.
- 98 sources on rebuttals were texts from the opponent's tradition. While summarising the opponent's point, the rebuttal cited the opponent's own book. On the results screen, sources are attached to that tradition's card, so Romans showed up on the Buddhism card and the Tao Te Ching on the Confucianism card. I fixed 32 rebuttals and added a rule to the checker that flags a source belonging to another tradition. An old manuscript file alone had 93 hits.
- Rebuttals piled up on Taoism. Across all 20 questions, Taoism was targeted 33 times and Sikhism 3 times, because the checker only looked at a cap within each question. I rewrote 21 rebuttals against new opponents to bring the range to 12 to 26.
- My own examples failed my own checker. The Buddhism example (141 characters) and the Christianity example (158 characters) in the instructions fell short of the 160-character minimum. The problem wasn't the writing; the minimum was too high. Getting the essence across in plain words naturally lands around there, so I lowered it to 120.
- Banned words caught normal words. The Korean for "Confucian school" matched inside the everyday word for "reason," and "Om" matched inside a common verb ending. I removed both.
- Rebuttals aimed at Confucianism and Sikhism showed
{{confucianism}}as is. The on-screen substitution rule only knew six traditions. I found it while running all 20 questions to the end. Now the rules are built from the list of traditions. - The final became "A vs. A." During the 0.28 seconds between matches, the previous match's button was still live, so a double-click overwrote the next slot with the winner. Now the buttons lock the moment you press.
I was told nothing showed up after tapping "In progress"
After deploying, the user said, "It's weird that nothing appears on screen when I tap In progress." The start button was at the top, but the bracket and the match appeared below the list of 20 questions. Where you'd tapped, all that was left was an "In progress" button that did nothing when pressed. I moved the start button next to the bracket, folded the question list behind "Change question," and made the screen move to where the match appears when you start. During a match the button becomes "Start over," and when it ends it becomes "Choose again," so you can restart at any time. Right after a restart, a timer the previous match had scheduled could make the new tournament skip ahead, so each tournament gets a number and old scheduled events are thrown away.
The design changed three times
At first it was an ink-colored background with a serif typeface. The user said it was "too AI-looking," so I switched to neo-brutalism, then, after receiving two reference images, to a bright grid with rounded cards, and then to what it is now: monochrome with scattered dots. The dots are drawn on a canvas. At first the density curve ran the wrong way and laid down an even noise; I fixed the curve so the dots scatter and fade out from one side.
Rebuilt with pop-ups, bottom to top
Along with the rename, the user asked to show matches in a pop-up so the one-on-one comparison feels stronger, to make the quarterfinal bracket climb from bottom to top, and to change the labels to A through H. A match opens in a full-screen pop-up (on mobile, a sheet that slides up from the bottom), and the results come up in a pop-up too. Closing the pop-up keeps your progress, and you come back via "Continue choosing." Then came "make the whole site feel like the new pop-up," so the card radius, shadows and A/B letter badges were applied across the page. The history went from horizontal bars to vertical bars with religion names on the x-axis, stacking one block for each choice. The terminology was unified as "choose" too.
After renaming it "Looking at Religion," I added material for "looking." Tap one of the eight tradition cards at the top for a short introduction: when it began, its key texts, its core ideas, and its branches. Pick a question and you can read all eight traditions' arguments, rebuttals and sources side by side with the identities revealed. The history keeps a table of the answers you were drawn to on each of the 20 questions. That table doesn't recommend or rule on any religion. Last, I added a table of similarities and differences. It boils down the eight traditions' answers on ultimate reality, what comes after death, the goal of life, and how they see people, one line each, and colors similar answers the same. That table is my own summary, so it can't capture every difference between branches.
How to use it
Tap "Start choosing" and the first quarterfinal opens in a pop-up. Read the two arguments, and if you like, mark which of the eight traditions you think each one is (each tradition can be attached to only one argument). Then tap "I choose this answer." When all seven matches are done, everything opens at once: the answer you chose all the way through, your guessing results, and all eight traditions' arguments, rebuttals and sources. Pick a question from the 20 under "Change question," or draw one at "Random." How many times you've finished and your guessing accuracy are saved only in this browser.
What I checked and what I couldn't
I ran all 20 questions × 7 matches = 140 matches automatically to the end, several times, and confirmed that no hidden words or unresolved markers were left. For the 525 citations from the Bible, the Quran, the Bhagavad Gita, the Tao Te Ching, the Dhammapada, the Analects and others, I checked that the chapter and verse numbers fall within ranges that exist, and no nonexistent numbers came up. At smartphone width (390 pixels), the bracket and match cards fit too.
What I couldn't check matters more. For the citations, a machine only checked that the numbers exist; no person has read each passage to confirm it supports the claim. The Sikh sources give only the book title, with no page numbers. The arguments were written by AI and are not the official position of any denomination, order, or individual believer. Most opening arguments are short, 120 to 145 characters, and since a different agent wrote each question, the same tradition's tone may vary a little from question to question. I'd encourage you to check the quoted passages yourself.