I write software for a living, which means I have a deeply unhealthy relationship with metrics and a genuine belief that anything you can count, you can fix. When I started learning French seriously, I applied the only methodology I trust: treat every mistake as a bug, log it, find the pattern, ship a fix. The problem was the logging. Reading a corrected version of my own text and trying to extract my own error patterns is like asking a developer to find their own bug by rereading the code that produced it. You wrote the bug. You can't see it.
Turning corrections into a defect log
The Correct My Writing and Explain Every Mistake Like a Real Teacher prompt produces exactly the artifact I needed: a structured table plus an error grouping with counts by type. That grouping is, functionally, a defect breakdown. Grammar: 6. Word order: 2. Register: 1. I started pasting those counts into a spreadsheet every week, one row per session, and within a month I had a chart of my own French falling apart in slow motion — and then, satisfyingly, repairing itself.
I run it on Gemini 3 Pro Preview because the table output is consistent enough to copy-paste straight into my tracker without reformatting, and because it's good at the synthesis step — when the same gender-agreement error shows up in three different sentences, it collapses it into one row with 'appeared 3 times' instead of making me notice the pattern myself. That collapsing is the part a tired human does badly at 10pm.
What the data actually told me
Two months of logs revealed something no single correction would have: my grammar errors were dropping but my register errors were flat. I kept writing emails that were slightly too casual for the relationship — tutoyer-ing people I should have vous-ed. I'd been treating it as a rounding error. The chart made it impossible to ignore, so I started running my actual outgoing French emails through my Write a Native-Sounding Business Email in Any Target Language prompt first, and the register mistakes finally started falling too. There is something genuinely clarifying about seeing your own incompetence as a bar chart. It depersonalizes it. The bar isn't a judgment about whether I'm a smart person; it's just a count, and counts go down when you work on them.
- The error grouping with counts is a defect breakdown you can chart over time
- Repeated errors get collapsed into one row with a count, which surfaces patterns a tired human misses
- Tracking by type, not total, revealed a flat category (register) hiding behind an improving one (grammar)
- Watching a real number drop week over week beats any gamified streak for actual motivation
- The 'why your native language causes this' column tells you whether a mistake is a typo or a systematic transfer error from English
The unexpected second-order effect
Once I'd been doing this for a while, I noticed the logging itself was changing how I wrote, before any correction happened. Because I knew every Tuesday I'd paste the week's writing in and watch the grammar count, I started writing more carefully during the week — not anxiously, just attentively, the way you write better commit messages when you know someone reviews them. The act of measuring changed the behavior being measured, which is normally a problem in data collection and here was the entire point. I was, in the most literal sense, debugging my own French by adding instrumentation.
I also started keeping the single best 'fixed pattern' from each week in a notes file — one line, the rule plus the example from my own writing. After a couple of months that file became the most useful French reference I own, because every entry is a mistake I personally made, in a sentence I personally wrote, which makes it impossible to forget in a way a grammar textbook never managed.
The honest caveat
This works because I feed it my own real writing every week — messages I actually sent, journal entries, the occasional unhinged review of a croissant. If you feed it textbook exercises, you get textbook results and your real-world errors stay invisible. The counts are only as honest as the input. The first few weeks I cheated by writing 'safe' sentences I knew were correct, and my numbers looked great and my French didn't improve, because I was logging clean code and ignoring the part of the app that actually crashed. Once I started feeding it the scary writing — the email I wasn't sure about, the message where I'd guessed at a verb — the counts went up at first, and that was the point where it started actually working.
If you like data more than you like vibes, grab the Correct My Writing and Explain Every Mistake Like a Real Teacher prompt on Prompt Dock, save the error grouping every week, and chart it. Then go fix your single biggest category. It's the most satisfying refactor you'll ever do, and the only one where the codebase is you.