I write issues and PRs in English most days, and Claude is open in the same terminal. Ask it to fix a sentence and it rewrites the whole paragraph. The text gets better. My English doesn't.
So I turned a repository into the opposite: Claude still reads everything I write, but it is told to point at the mistake and wait for me to fix it, then write the mistake down.
- Source: https://github.com/0-draft/claude-english-coach
- Use: Use this template on GitHub, clone your copy, run
claudeinside
What's in it
CLAUDE.md rules Claude follows in this repo
phrases.md mistake → correction, with a count
vocab.md words you didn't know, with an example
journal/YYYY/MM-DD.md what you wrote each day
tech/README.md reusable phrases for issues, PRs, review replies
app/rounds.html a drill app (below)
docs/research-ja.md the research the rules are built on
The coach itself is just CLAUDE.md: nothing to install. phrases.md and vocab.md start as empty tables, journal/ as an empty directory, and they fill up as you write.
The first line to edit is the native language, because feedback is given in it:
Native language: **Japanese** (change this line to your own language)
The loop
1. Write something in English. A chat message, a diary entry, a draft PR description.
2. Claude answers the content first, then marks errors. It says where the error is and what kind it is (article, tense, preposition...), and does not give the corrected sentence. You try again. Only if the second attempt is still wrong does it show the answer and the rule. From CLAUDE.md:
- Prompt before you correct. Mark where each error is and what kind it is (for example `article`, `tense`), and let the user try to fix it. Give the corrected form and the rule only if their second attempt is still wrong. Prompts that make learners self-correct outperform simply recasting the sentence (Lyster & Saito 2010).
- Keep corrections minimal. Fix what is wrong or clearly unnatural; do not rewrite style or rephrase sentences that already work. LLMs tend to over-correct, and that hides which errors actually matter.
- Never write the user's text for them before they have tried. If they ask "how do I say X", ask for their attempt first.
If you write in your native language, it asks you to try English first.
3. Claude logs it. The exchange goes into journal/YYYY/MM-DD.md, split into what you fixed yourself and what needed the answer. Each mistake becomes a row in phrases.md:
| Mistake | Correction | Category | Note | Count |
| --- | --- | --- | --- | --- |
| I will fix it until tomorrow. | I'll fix it by tomorrow. | preposition | deadline is "by" | 1 |
Making the same mistake again bumps Count instead of adding a row. New words go to vocab.md, always with an example sentence. At the end of the session Claude commits with docs(journal): YYYY-MM-DD, so git log is the study history.
4. Pick a focus. phrases.md has a ## Focus section for one or two categories a week, chosen by the highest counts. Errors in those categories get full feedback. Everything else is mentioned only if it blocks understanding.
5. Say quiz. Claude asks five questions from phrases.md and vocab.md, weighted toward high counts and the focus categories, one at a time. They are production questions: a meaning or a sentence with a blank, and you type the English. No multiple choice.
Why these rules
Each rule in CLAUDE.md names the paper it comes from, and docs/research-ja.md (Japanese) has the sources. The three that shaped the loop:
- Prompting learners to fix their own errors beat recasting the sentence for them (Lyster & Saito 2010).
- Students who had free access to GPT scored 17% lower than the control group once it was taken away (Bastani et al. 2025, a math RCT).
- ChatGPT tends to rewrite parts that were not wrong (Fang 2023), which hides the errors that matter.
The drill app
The Markdown loop needs a session of writing. For short daily practice there is app/rounds.html: one HTML file, made to run as a Claude artifact. Opened in a plain browser it works too and saves to localStorage. The UI is in Japanese.
- Review: spaced repetition with ts-fsrs 5.4.2. Every card makes you type the English.
- Listen: dictation one sentence at a time. Missed words become cards. Playback speed adjusts to keep you at 80-90% of words right.
- Speak: role-play a stand-up or an incident report with Claude, or the 4/3/2 drill (same talk in 4, 3, then 2 minutes). It records how many words you spoke.
The cards come from the same files. Paste the phrases.md or vocab.md table into the import box and each row becomes a card. I pulled the parser out of the page and ran it on a small table I wrote (output trimmed):
$ node run.js app/rounds.html table.md
[
{
"kind": "fix",
"prompt": "I fixed the bug in yesterday.",
"answer": "I fixed the bug yesterday.",
"category": "preposition",
"note": "no \"in\" before yesterday"
},
...
{
"kind": "vocab",
"prompt": "不安定な (テストが)",
"answer": "flaky",
"example": "The test is ____ on CI.",
"note": ""
}
]
A vocab row gets its word blanked out of the example sentence, so the card asks for the word in context.
The habit grid runs for 66 days, the median time to form a habit in Lally et al. (2010). Five minutes keeps the streak, 15 meets the goal, and one missed day a week is forgiven. docs/apps-research-ja.md goes through which Duolingo mechanics made it in and which didn't. There are no XP, leagues, gems or hearts.
The artifact viewer sometimes decodes the page in a charset other than UTF-8, and every Japanese string turned into mojibake. app/build.py escapes every non-ASCII character in the HTML, CSS and JS so it can't happen:
$ python3 app/build.py app/rounds.html dist/rounds.html
$ LC_ALL=C grep -c $'[^\t -~]' app/rounds.html dist/rounds.html
app/rounds.html:292
dist/rounds.html:0
The ASCII copy is the one you publish, with the db, user and sample capabilities. sample is what lets the page call Claude. Without it the role-play and the "why?" explanation buttons stay hidden. Artifact frames block the microphone, so speech goes in through OS dictation (fn twice on macOS).


Top comments (0)