Guides

Everything we have written down

Replacing an expensive model with a cheaper one is only worth doing if you can prove quality held. These are the pieces of that argument: how a prompt gets optimized, what makes a set of examples trustworthy, why prompts decay, and how to move off a platform that is closing. Every number in them comes from a run you can repeat.

Making a prompt better, provably

Start here if you have a prompt in production and no number that says whether it works.

Prompt optimization that proves itselfProductHow GEPA rewrites a prompt against your verified examples, and why the score that matters comes from rows the optimizer never saw.July 12, 2026
Your prompt worked. Then it got worse.GuideThe four reasons prompts decay in production, why you find out last, and the six-step regression test that catches it before your users do.July 15, 2026
Receipt extraction, eight models, one held-out sliceBenchmarkOur public benchmark in full: 200 real scanned receipts, every score reproducible, including the runs that did not go our way.July 12, 2026
The last hand-tuned promptEssayFor the people who fund AI work rather than write it: what changed when prompt tuning stopped being a person guessing, in plain English.July 16, 2026

Deciding what to adopt

Start here if you are choosing between tools, surfaces, or a way off a platform that is closing.

Or skip the reading and run the loop on one of your own tasks. Start a free pilot, or read the quickstart.