Metadata-Version: 2.5
Name: chat-export-insights
Version: 0.1.1
Summary: See what you actually use ChatGPT and Claude for. Runs on your computer; your conversations never leave it.
Requires-Python: >=3.10
Description-Content-Type: text/markdown

# chat-export-insights

See what you actually use ChatGPT and Claude for, from your own data export.
**It runs on your computer. Your conversations never leave it:** no uploads, no AI model reads them, no network access at all.

## Run it

1. Export your data:
   - ChatGPT: Settings → Data controls → Export data.
   - Claude: Settings → Privacy → Export data.

   You'll get a `.zip` by email.
2. Install [uv](https://docs.astral.sh/uv/) once:
   - macOS/Linux: `curl -LsSf https://astral.sh/uv/install.sh | sh`
   - Windows: `powershell -c "irm https://astral.sh/uv/install.ps1 | iex"`
3. Run:

```
uvx chat-export-insights ~/Downloads/your-export.zip
```

It writes `findings.md` (for you) and `findings.json` (for tools) next to the zip. Then it offers to delete the zip.

## What you get

- The kinds of work you use AI for (email, summaries, research, proposals, code… 22 kinds, 17 languages), with how often per week.
- Prompts you keep retyping, shown as templates where anything that varies (names, companies, numbers) is replaced by `[X]`.
- When you use AI (hours, weekdays) and in which languages.

## Privacy

- No network code: the tests fail if the package imports any networking module or opens a connection.
- Output contains counts and keywords, never your messages or conversation titles.
- Templates keep only words that recur across several prompts. **Read `findings.md` before you share it.** Use `--no-snippets` to leave templates and unmatched words out entirely.
- The only files written are `findings.md` and `findings.json`. The export is deleted only if you confirm at the prompt (Enter or y); `n`, Ctrl-D or Ctrl-C keep it. Use `--keep-export` to skip the question.

## Options

```
chat-export-insights EXPORT [--out DIR] [--no-snippets] [--keep-export] [--version]
```

Keyword lists for most languages are machine-drafted and not yet reviewed by native speakers. The report says which ones.
