Which AI Writes the Best Lesson? I Made Them Grade Each Other

More from Building MarrowMinded — the reading app I build with Claude Code. Start here: I built MarrowMinded.

MarrowMinded gives you one lesson a day from a book. Under the hood it can run on different AI models — Claude, GPT-4o, Gemini, and Groq’s Llama. Which one writes the best lesson? I’d been guessing. So I built something to actually find out.

The setup

Same chapter, four models, one question: whose lesson is best?

The trick is who judges. A single judge has favourites. So I rotated it: each model grades the other three, and every lesson gets scored by three different judges. Then one more pass where Claude grades all four — including itself — to check whether a model plays favourites. A small script adds it up and logs every score to a table.

I built the whole thing in n8n — four model nodes, the judges, the scoreboard — so I could see each part instead of burying it in code.

The result wasn’t clear-cut — which is the interesting part

By average score, Claude won: 9.33. Gemini close behind at 9.0, GPT-4o 7.67, Llama 6.67.

But by head-to-head wins, Gemini took three of the four rounds. Claude scores highest on average; Gemini gets picked as the single best most often. Two honest ways to count, two different winners. A real test shows you that instead of smoothing it over.

Does Claude favour Claude?

When Claude graded all four, it picked itself. Suspicious. But two things clear it: the three other judges also ranked Claude top, and when Claude judged a group that didn’t include itself, it picked Gemini. So it’s not playing favourites — it just rated the best lesson best.

That check — does the judge favour its own kind? — is the part I’m proudest of. It’s easy to crown a winner. It’s harder to show the crown is fair.

Two surprises

• Gemini’s small “flash” model punched way above its weight. The cheapest and fastest of the four, and it nearly won.
• Llama came last as a writer — and was the worst judge too. It sometimes returned broken output and scored the strongest lesson a 1. The test grades the judges, not just the writers.

The hard part wasn’t the AI

Honestly? The models were the easy bit. The two hours of pain were billing and paperwork: one provider out of credits, another that had quietly retired the model I was calling, and a third where I’d topped up the wrong account entirely. None of it was the actual work. When you wire up four vendors, most of your time goes to their plumbing, not yours.

Takeaways

• Don’t trust one judge. Rotate it and average across several. One grader has taste; three have a signal.
• Test the judge, not just the answer. Check whether it favours its own kind before you believe the ranking.
• Count more than one way. Average and head-to-head disagreed here — both were true.
• Budget for vendor plumbing. The AI is a box; the accounts around it are where the day goes.

If you had to bet — which AI writes the best lesson? Come read one and judge for yourself. → olekwrites.com/marrowminded