Your AI Should Grade You Back

Everyone is trying to make their AI smarter. A better model, a bigger prompt library, a folder of clever system messages copied off the internet. Almost nobody stops to ask the more uncomfortable question: am I any good at driving this thing?

I use Claude Code every day. It writes, refactors, ships. And like most people, I assumed that when a session went sideways it was the model's fault. The model got confused. The model lost the thread. Then I read back through my own prompts and saw the truth. Half the time the model was confused because I was.

Here's the part nobody tells you. Every prompt you have ever typed is sitting on your machine. Claude Code keeps a full transcript of every session. That is not a privacy problem to be scared of. It is a mirror. It is the most honest record you will ever have of how you actually think under pressure: what you asked for, how you asked, where you went vague, how many times you had to double back and correct.

So I built a tool to read it. Not a product, not a launch. Something for me: a way to see my own habits clearly and chip away at the worst ones. I still tweak it most weeks.

What it does

It is a Claude Code skill, a small self-contained capability you can drop into the tool. It runs entirely on your machine, reads only, and sends nothing anywhere. It scans your own history and does four things.

It scores your prompting out of 100, with a breakdown: how specific you are, whether you give instructions or vague nudges, whether you work in a plan-then-approve rhythm, and whether you have built any reusable tooling for yourself at all.

It names your anti-patterns and counts them. Not "be more specific." The actual ones. The exact number of times you typed "still broken" instead of saying what was broken.

It quotes your worst prompts back at you and rewrites them in your own voice.

And then, the part I care about most, it does not stop at advice. It offers to build the fix: the custom commands and the sub-agents for the workflows you repeat forty times a week.

The rewrite that stung

The most useful output was not the score. It was the before and after. It takes something you actually typed and shows you the sharper version.

One of mine was, roughly, "that's still not working, try again."

It rewrote it to: "The build fails with this error. You changed X in the last step. Revert that, try Y instead, then run the tests and show me the output." Same intent. One version wastes a round trip and a chunk of context while the model guesses what "not working" means. The other gets it right the first time. I do the lazy version more often than I would like to admit, usually when I am tired, which is exactly when I can least afford the wasted loop.

Then it built me things. It looked at the verbs I type most, found that I run the same verify, commit, push dance dozens of times a week, and offered to fold it into a single command. I said yes. That is the difference between a report and a tool. A report tells you that you are repeating yourself. A tool stops you having to.

The finding I did not expect

It also scans for secrets. The first time I ran it on my own history it flagged old credentials I had pasted into a chat and forgotten about, sitting in plain text on my disk ever since. That is not a Claude problem. That is a me problem, and it is exactly the kind of thing you only find when you finally look at what you have been doing.

Why the coach, not just a better model

The models keep getting better on their own. My prompting was not. The ceiling keeps rising, and it is easy to keep typing the same lazy one-liners you typed a year ago without noticing the gap widen. I certainly was.

When the tool is this capable, the constraint moves. It stops being the model and starts being the person holding it. Leverage is no longer about access to AI. Everyone has that now. It is about how well you can direct it.

I keep coming back to what happens to work when automation gets cheap. This is the same argument at desk scale. When the machine will do almost anything you can clearly specify, the scarce skill is clear specification. The bottleneck is you. And the only way to fix a bottleneck is to measure it.

For the record, it gave me a number, and I'm sure plenty of people score higher. That was never the useful part. What mattered was what came next: it named exactly where I fall short. I over-ask when I should instruct, and I under-build reusable tooling for work I repeat constantly. Reading that back from a script parsing my own words was more useful than any prompting guide I have read, because it was not generic. It was me.

We spend all our attention grading the AI. It is worth spending some of it letting the AI grade us back. The data is already sitting there. I just had to be willing to look at it.

prompt-coach is on GitHub if you want to run it against your own history.