Proven, Better, New

Decide what to build before you build it. Your AI researches what already exists, checks whether your version is actually better, and keeps the risky new part small.

1. ProvenDoes this already exist, and does it work? Who tried and stalled? If something is good enough, just use it.
2. BetterEach improvement checked against a named example: clearly better, about the same, worse, or test it first.
3. NewOnly what nobody has done. About 20% of the product, at most: it's the risky bet.

At the end you get a verdict and a build brief to paste into Lovable, Cursor, Claude Code or Bolt: what to copy, what to improve, and the small new part, marked so it can be cut if it fails.

Proven, Better, New is Mark Pincus's framework, from his years building Zynga and his book Life at the Speed of Play. All credit for the method is his. This free tool puts it to work in your AI; it isn't made by or affiliated with him.

Add it

Claude (web, desktop, phone) Settings → Connectors → Add custom connector. Name it "Proven, Better, New" and paste: https://pbn.builditbox.net/mcp Then, in a chat: "Use Proven, Better, New on my idea: …"
Claude Code claude mcp add --transport http pbn https://pbn.builditbox.net/mcp
Anything else that speaks MCP Add a remote (streamable HTTP) server at https://pbn.builditbox.net/mcp. No sign-in.

What it keeps

Your idea stays yours. The thinking happens in your own AI chat; this server gives it the method and the checks, and never stores what you send.

To count who has already tried your idea, the server searches GitHub and Hacker News with a few keywords your AI picks. Those keywords go to their search; the server keeps nothing.

At the end your AI offers to save, and asks first. Saving records only a nickname you choose, the date, the verdict and how many items were in Better and New. Never the idea, its title or its text. That's how we learn whether people come back with a second idea. Say no and nothing is kept.

Made at builditbox.net.