Learn a Language While Coding: Turning Your Workday Into Vocabulary You Keep
A developer learning English is in an odd position. The language is not a subject waiting for you in the evening; it is the medium you work in for eight hours. Error messages, library docs, review comments, the half-understood answer on Stack Overflow. That is a lot of real input at roughly the right level, and almost none of it survives the week.
Nothing is wrong with your memory. The word showed up in a stack trace, you worked out roughly what it meant from context, you shipped the fix, and by Thursday it was gone. Understanding a word when it is in front of you and being able to produce it in a meeting are two different skills, and a workday full of reading only ever trains the first one. The material is excellent. The loop is broken at the end.
The words in your workday beat the words in an app
A course teaches you the vocabulary that is frequent in general English. Your job repeats a narrower and far more useful set: deprecated, idempotent, throttle, fallback, brittle, workaround, regression, edge case. Those are not exotic words to a native speaker, and they are exactly the ones you will need again tomorrow, because your work generates them on a loop.
The more interesting half is the connective language around the code rather than the jargon inside it. Softening a disagreement in a pull request, saying that an approach feels risky without saying it is wrong, writing a message that asks for something without sounding like a demand. That register is hard to get from a textbook and impossible to avoid at work, and it is sitting in your review threads for free.
Why the obvious plan collapses
Everyone has tried the same thing: a notes file, or a scratch tab, where the interesting words go. It fails in two predictable places.
The first is the moment of capture. You are three levels deep in a debugging session, holding the shape of the problem in your head, and adding a word means leaving the editor, finding the app, typing a translation, choosing a deck. That is expensive at exactly the wrong moment, so you promise yourself you will do it later, and later is a lie you tell about twenty words a week.
The second is what happens after. Suppose you do capture them. A list is not learning. Without something bringing each word back before it fades, the file becomes an archive of words you once found interesting. Capture without a schedule produces a very tidy record of everything you forgot.
What actually changed: where the assistant sits
You almost certainly have an AI assistant open inside your editor already, in Cursor or Claude Code or Zed or Continue. When a word or a phrasing puzzles you, that chat is where the question goes. It is the one place in the workflow that already knows what you are reading, and it costs nothing to talk to because you are talking to it anyway.
The piece that was missing until recently is memory. A chat explains the word beautifully and then forgets you ever asked. The Model Context Protocol, MCP, is the standard that lets an assistant call outside tools, which means it can now write to something that persists rather than answering into the void. Connect a vocabulary tool over MCP and the capture step stops being a context switch: you ask what "brittle" means in that review comment, and the word lands in a review queue as a side effect of the question you were going to ask regardless.
That is a small mechanical difference with a large practical one. The cost of saving a word drops to zero, and the words you save are the ones you personally stumbled on, which is a much better filter than any frequency list.
| Notes file plus a separate app | Assistant connected to your deck | |
|---|---|---|
| Where capture happens | Another window | The chat you already opened |
| Cost at the moment of capture | A context switch | None, it is a side effect |
| What decides which words go in | Whatever you remember later | What actually confused you |
| What happens next | Usually nothing | It comes back on a schedule |
| Typical failure | You stop adding by week two | You skip review days |
Setting it up so it survives a real week
The setup matters less than the standing instruction. An assistant will not capture anything unless you have told it, once, that this is part of the job.
- Connect the vocabulary server to your client once, then forget it exists.
- Give the assistant a standing rule in your project instructions: when I ask about an English word or phrasing, save it, and do not interrupt what we were doing to tell me.
- Ask about the phrasings, not only the words. "Is this comment too blunt" teaches you more than a definition does.
- Once a week, write something in your target language and ask for corrections on the parts you were unsure about.
- Review on your phone, away from the machine. Five minutes on the tram is enough, and it is the half that most people skip.
The part that is easy to oversell
This is a capture and retention loop, and it is worth being blunt about what it does not touch. Speaking is untouched. Nothing about reading docs and typing in a chat window trains the muscle you need when someone asks you an open question in a standup, and no amount of clever tooling substitutes for producing sentences out loud with another human.
Technical English is also a narrow register. It will make you fluent in your job and leave you stranded at a dinner party, so the wider reading and listening still has to happen somewhere. And the loop only pays if you actually meet the words again; an assistant that captures a hundred words a week and a review habit of zero is worse than useless, because it feels like progress.
If English is not the language you are learning
The same loop runs the other way, with one adjustment: your workday will not supply the input, so you have to. Tell the assistant to answer in Spanish or German or Japanese, keep the code in English, and let the explanation carry the language. Any word you have to stop and query gets captured on the spot, and the difficulty tunes itself to whatever you happened to be reading.
Where MindDory fits
MindDory is a vocabulary brain your assistant can reach. It runs as a remote MCP server, so Claude, Cursor and other MCP clients can read your level and your due cards, capture a word or a grammar mistake straight out of a conversation, and mark a word as demonstrated when you used it correctly. The full tool list and the client setup are on the MCP language learning page.
What happens to the word afterwards is the point. Each captured word becomes a card with a memory cue, and your sessions are built from the words closest to slipping rather than whatever you added most recently, which is the spaced repetition idea applied to a deck you never had to build. It syncs to web, iOS and Android, so the reviewing happens away from the desk where the capturing happened. If you want the broader version of this argument, the guide on learning English with AI covers the conversation side.
You are already spending the hours. The only question is whether the language passing through them leaves anything behind.