Agora

Inspiration

We wanted to know if you could digitize how a person reasons — not their conclusions, which an LLM can parrot, but the actual generator: their priors, their forbidden moves, their blind spots. If you could, you could run them forward on questions they never answered. That felt worth building.

What it does

Two halves, one primitive — a worldview file.

Worldview: a guided capture flow that infers how you reason by watching you work through real cases, then exports it as a portable markdown file any agent can run as you. The Arena: put two brains in conversation. Pick from ten historical figures (Nietzsche, Socrates, Aristotle, Napoleon…) or your own worldview. Watch them debate with a structured five-phase format, or step in and spar yourself.

How we built it

React + Vite frontend on Cloudflare Pages, Hono Worker backend, Claude for generation. The key architectural choice: a deterministic state machine owns turn structure and phases — LLMs only generate utterances. Predictable, replayable, debuggable.

Worldview files are eight ordered sections (Identity, Epistemology, Core commitments, Forbidden moves, Rhetorical signature, Source texts, Blind spots, Anachronism policy). Order matters — LLMs weight earlier content more heavily.

Challenges we ran into

Fighting "Claude in a costume." LLMs default to balanced and agreeable. Three things counter it: Forbidden moves that block the agreeable response, Quote anchors that give concrete voice texture, and Blind spots that force specific failure modes. A worldview without failure modes isn't a person.

Avoiding false balance. Cross-examination targets the weakest point. The moderator names sub-point winners and flags unanswered objections — never picks an overall winner.

Accomplishments that we're proud of

The store generators, not conclusions principle holds in practice. Watching Nietzsche and Confucius debate the nature of hierarchy, each feeling distinctly themselves across every turn — that was the moment we knew it worked.

What we learned

People are unreliable narrators of their own reasoning. The capture flow has to watch someone reason through real cases — asking "how do you think?" produces flattering noise, not the actual generator.

What's next for Agora

Gonna get some users. Shareable worldview libraries. Bookmarkable moments as the viral unit. Retrieval grounding to flag hallucinated quotes. Worldview archetypes ("a contemporary libertarian") so you can test your brain against a position, not just a person.

Built With

• anthropic
• claude
• cloudflare-pages
• cloudflare-workers
• hono
• react
• tailwind-css
• typescript
• vite