Find the English mistakes you actually make in Claude Code.

You already write English in Claude Code every day, and nobody ever corrects it. The audit reads that history back from your machine and lists the mistakes you keep making. Nothing to submit, no sample to write.

Your messages never go to Glite. Only the list of mistakes does, and only after you see every line and choose to send it.

See how to run it View the repository
glite audit
illustrative run · local review

You: /audit

English audit38% complete
Step 4 of 5
Finding English mistakes
Claude Code
205 of 512 messages processed — 40%
Collected so far
512 eligible messages · 14,900 English words
Estimated time
42–115 minutes
Mistakes to include12
Example only. The review page opens on your own machine with every shareable mistake ticked — untick anything you would rather keep. Glite gets the list; nobody new gets your messages.
Only Claude Code It reads one source: the history of the tool you are already using.
You see the list first Nothing is sent before you have read it and chosen.
Glite never sees your messages Only the list of mistakes, with personal details removed.

Run it in Claude Code.

You need Python 3.12 or newer, uv, and Claude Code. Run these commands in your terminal so Claude starts inside the cloned repository and can load the audit skill.

Clone, install, and launch

git clone https://github.com/GliteTech/glite-english-audit.git
cd glite-english-audit
uv sync --locked --all-groups
claude

Run the audit skill

When Claude Code opens, enter the project skill command:

/audit

Or ask in plain English Run an English audit.

Using Codex? Open the same cloned directory in Codex and enter Run an English audit. It reads your Codex history instead.

What it does.

It asks you once, then runs on its own. You can walk away.

  1. Reads your Claude Code history

    Local scripts count what is there. No message text goes to the model during this.

  2. Finds the mistakes

    Claude Code reads the messages you typed into it and flags the constructions that repeat.

  3. Shows you the list

    A local page lists every mistake. Exclude anything you would rather not send.

  4. You send the list, not your messages

    Glite receives the mistakes with personal details already removed, and returns a full report.

A report based on your real English.

The report shows what repeats in your submitted records. It does not assign a CEFR level or pretend to measure all of your English.

Glite English Auditillustrative report

Your most frequent problems involve question structure, determiners, and relative clauses.

Verified mistakes per 1,000 analyzed words 4.6
  1. 01Question and clause word order4 records
  2. 02Articles and determiners3 records
  3. 03Relative clause words2 records

Based on 16 verified mistakes across 3,449 analyzed English words. Example figures only.

  • Patterns, not a generic score

    See which kinds of mistakes repeat and how many submitted records support each pattern.

  • An honest rate

    Counts use the amount of English actually analyzed, with withheld mistakes included in the overall rate.

  • Coverage and limits

    The report states what was checked, what was withheld, and what the audit cannot conclude.

Your raw messages do not go to Glite.

This is local-first, not 100% local. Local scripts find your English on your computer. No AI reads your messages during that step. Text you explicitly select is analyzed through your active AI provider. Glite receives only the privacy-safe records you review and approve.

  1. 01 · Your device Find and select

    Local scripts find candidate English. You choose the sources, period, and what proceeds.

  2. 02 · Your AI provider Analyze selected text

    Claude Code reads the messages you typed into it. No new party sees them. Glite is not part of this.

  3. 03 · Glite Build the report

    Only reviewed, privacy-safe mistake records and anonymous counts are submitted if you choose.

Questions people ask first.

Does Glite see my messages?
No. It receives a list of mistakes — the wrong construction, the rule it breaks, and a short example. You read every line before you send it, and you can exclude any of them.
Is there a privacy risk in letting Claude Code read this?
No new one. These are messages you typed into Claude Code, so it already received them when you wrote them. Reading them back discloses them to nobody who did not have them.
Why only one source?
Because it is enough. Your Claude Code history holds more natural English than any test would, and reading one source keeps the whole thing explainable in three sentences. If you turn out to have too little there, the audit offers to read Codex, Cursor or your dictation history as well.
What if I use Codex, not Claude Code?
No problem — do exactly the same thing. Open the cloned directory in Codex and ask it to run an English audit. It reads your Codex history rather than your Claude Code history, and says Codex throughout. The audit always reads the history of the agent it is running inside, and never the other one.
How long does it take?
About 20 to 40 minutes. It runs unattended, so you can start it and do something else.
What is a good report?
One that shows a habit repeating, not a one-off. That takes a few tens of thousands of your words — roughly two weeks of ordinary use. The audit works the period out for you and asks once before it starts. How many mistakes that turns into depends on you, so it does not guess.

Ready to see your patterns?

Clone the repository, launch Claude Code inside it, and run the project skill.

Run the audit View the repository
Already have a submission package? Upload it here.

Create your report manually.

Choose the package file you downloaded, or paste it below. A normal audit opens this page for you, so most people do not need this form.

or paste the JSON below

Privacy Policy · Terms and Conditions

The report opens on this website. Submissions are kept under the Terms linked above.