English | 中文
Oral English Practice turns the Claude app into a tireless speaking partner and Claude Code into the coach that remembers everything. You practice out loud — real conversation and listening — and every session comes back as a graded report, a verbatim record of the words you stumbled on, and a clear focus for next time. Over weeks, the fleeting act of "talking to an AI" becomes a measurable climb toward native-level fluency: your scores trend, your recurring mistakes get hunted down one by one, and the difficulty quietly adjusts to keep you at your edge.
The trick is splitting the work across the two tools that each do half the job. The app has a voice but no memory; Claude Code has memory but no voice. This skill is the bridge — you carry two short pastes between them, and it handles the rest: archiving, scoring, error-banking, trend charts, and the next plan.
One subscription, two pastes, long-term practice.
1. One subscription, two pastes, long-term practice. No new app, no extra account, no fee beyond the Claude subscription you already have. Each round is just two copy-pastes — the prompt out, the report back — yet they compound: a few minutes of speaking becomes weeks of tracked progress instead of practice you forget by tomorrow.
2. A graded report after every session. The moment you finish, you get a clear scorecard: an estimated CEFR / IELTS level, a "distance to native /100", and seven 1–10 dimensions — fluency, lexis, grammar, pronunciation, discourse, interaction, listening. Progress becomes something you can see, not just feel.
3. A mistake bank that hunts your recurring errors until they're gone. Every stumble, grammar slip, and unnatural phrase is logged verbatim. Patterns that keep coming back are confirmed, counted, and re-drilled each session — and retired only once you go clean.
- IELTS / TOEFL / PTE candidates — the speaking section is the hardest part to practice alone. Here you get unlimited spoken practice with a graded report and targeted drills after every session.
- Anyone living, working, or studying in an English-speaking country — build the confidence to sound natural in meetings, interviews, classes, and everyday conversation.
- Professionals who work in English — rehearse presentations, client calls, and interviews, and steadily close the gap between "correct" and "native".
- Self-driven learners aiming for native-level fluency — turn daily speaking into a measurable, compounding climb instead of practice you forget by tomorrow.
- 🎙️ Ready-made coaching prompt — paste it into the Claude app and start a native-level voice session (conversation + listening) right away.
- 📊 Scored assessment report every session: a CEFR/IELTS estimate, a "distance to native /100", and seven 1–10 dimensions.
- 📈 Long-term progress tracking — every session logged to
data.csv, with trend charts (or a text trend if you have no Python). - 📝 Verbatim error transcript — every stumble, grammar slip, vocab gap, and unnatural phrase tagged inline and explained.
- 🏦 Mistake bank — recurring errors are confirmed, counted, and re-drilled until you go clean, then archived.
- 🎚️ Adaptive difficulty — the next-session focus tells the coach what to ease or push, based on your actual scores.
- 🎭 Fully customizable — change the coach's tone, topics, and correction style in plain language (or by editing one file).
- 🌐 Bilingual & language-adaptive — English / 中文 docs, and the skill replies in whatever language you write in.
- 🔒 Local-first & private — your records stay on your machine; no uploads, telemetry, or third parties.
- 🪶 Minimal setup — just Claude Code + the Claude app; Python/matplotlib is optional (only for chart images).
Claude app (practice, stateless) Claude Code + this skill (long-term brain)
· voice conversation + listening · parse report, archive
· outputs: annotated transcript → · append to data.csv, chart trends
+ report + DATA BLOCK · find stubborn weaknesses
· write next-focus
← paste next-focus into the app ─────── · continuity across sessions
- Get the prompt — in Claude Code, run
/oral-english-practice(or say "I want to practice spoken English"). It hands you theapp-promptplus last session's carry-over focus. - Practice — paste the prompt into the Claude app (voice mode), talk for 15-30 min. The coach ends with an annotated transcript + a scored report + a one-line DATA BLOCK.
- Log it — paste that whole output back into Claude Code. The skill files it, updates your trend data, and writes your next focus.
- Review — every few sessions, ask for a trend review/chart.
/plugin marketplace add XcutingUnicorn235/Oral-English-Practice
/plugin install oral-english-practice@xcutingunicorn235
One command, plus automatic updates (/plugin marketplace update).
Windows: double-click install.bat.
macOS / Linux: bash install.sh.
By hand: copy plugins/oral-english-practice/skills/oral-english-practice/
into ~/.claude/skills/.
Re-running the installer is safe — it never touches your data.
Required — nothing beyond these two:
- Claude Code — runs the skill (logging, mistake bank, backups, diagnosis).
- Claude app — voice practice + listening.
That's all you need. Every core feature works with just these — the skill uses Claude's built-in file tools, and even backups work without Python (Claude copies the files itself).
Optional:
- Python 3 +
matplotlib— only for the visual PNG trend chart. Without it, the trend review still works, just as a text/markdown summary. You can add it anytime later withpip install matplotlib.
Created on first use, in a clearly named folder so you always know where your records are. Default location:
- Windows:
%USERPROFILE%\oral-english-practice-log\ - macOS / Linux:
~/oral-english-practice-log/
The skill prints the full path the first time it creates the folder.
You can keep your records anywhere you like. Two ways to change the location:
- Easiest: just tell your Claude, e.g. "move my practice log to
D:\English\" — it sets it up and moves any existing records over. - Manual: create a file
~/.oral-english-data-path(in your home folder, a dot-file with no extension) whose first line is the path you want, e.g.D:\oral-english-practice-log. The skill uses that path from then on.
| File | Purpose |
|---|---|
app-prompt.md |
The prompt you paste into the Claude app (coach persona + topic pool + report format) |
data.csv |
One scored row per session — your trend dataset |
transcripts.md |
Verbatim error log; every stumble/error tagged inline and annotated |
mistakes.md |
Mistake bank — confirmed recurring errors with frequency + status, re-drilled until fixed |
sessions/ |
Full report per session (session-NN-YYYY-MM-DD.md) |
next-focus.md |
Carry-over card; includes a block to paste into the app next time |
backups/ |
Timestamped snapshots taken before each logging write (the optional Python helper prunes them to the last 10) |
session,date,cefr,ielts,native,fluency,lexis,grammar,pronunciation,discourse,interaction,listening
session is the integer session number (the primary key). Scores are 1-10 per
dimension; native is distance-to-native out of 100; a dimension that couldn't
be judged (e.g. pronunciation in a text-only session) is recorded as NA.
Logs from v1.x (no session column) are migrated automatically on first use.
| Tag | Meaning |
|---|---|
| [G] | Grammar (tense, articles, agreement, prepositions...) |
| [V] | Vocabulary gap — couldn't find / used the wrong word |
| [N] | Unnatural / not idiomatic |
| [F] | Fluency stumble — filler, false start, long pause |
| [P] | Pronunciation |
- Your records stay on your computer. The skill has no network code — no
uploads, no telemetry, no analytics, no cloud sync. Your data files are
written only to your local data folder, and
.gitignorekeeps them out of git. - But the skill runs through Claude, which is a cloud AI. Using it means talking to Claude Code and the Claude app, so the text you process — your reports and your spoken practice — passes through Anthropic's servers as part of normal Claude usage, like any Claude conversation. That's inherent to using an AI model, not the skill sending data anywhere extra.
- No third parties. Nothing goes anywhere beyond the Anthropic services you already use by running Claude; there is no separate cloud copy of your records.
- Coach tone / topics / correction style: edit the prompt block in your
app-prompt.md, or just ask your Claude to change it in plain language. - Scored dimensions: if you change them, keep the DATA BLOCK line and the
data.csvheader in sync, or the skill can't parse new sessions.
Just point a friend at this repo: they install the plugin (the two /plugin
commands above), or drop
plugins/oral-english-practice/skills/oral-english-practice/ into their
~/.claude/skills/. Their data is created fresh on their own machine — the
skill is the engine, the data is personal — so nothing of yours travels with it.
Built following Claude's official skill-creator best practices and adapting
patterns from the fluent language-learning kit — a mistake bank,
write-before-write backups, and an adaptive-difficulty signal — while staying
deliberately lightweight (CSV + Markdown, no databases) for a single-user,
native-level goal.
See CHANGELOG.md for the version history. Latest: v1.1.0 — data-integrity hardening (Python-free backups, validated logging, schema-versioned data, smarter mistake bank).
Released under the MIT License © 2026 XcutingUnicorn235.