Prose Trainer
Browser tool that inspects recurring writing patterns in a draft and compares two drafts side-by-side. Analysis runs locally; nothing is uploaded. Findings are prompts for review, not evidence of AI authorship.
What it does
Paste a draft. The tool inspects recurring writing patterns — three-item parallel clauses, hedge-affirm pairs, and a small set of other stylistic shapes it knows about. Click a finding and the matched span focuses in the draft. Scan analyses one draft; Compare pastes two drafts side-by-side and shows what changed.
A match gives you something to look at. It does not tell you the writing is bad, that an AI wrote it, or that the line should be cut. Findings are prompts for review, not verdicts.
What exists today
A browser-only tool at /lab/prose-trainer/. No backend, no upload, no account. Draft analysis happens in your browser.
Demonstration
Six small cases show what the tool actually does on synthetic input. They were not all designed at once. Cases 1–5 were defined first; case 6 was added after the case-5 expectation turned out to be wrong.
One input is reproduced below for you to try. The table summarises all six cases, which used short, made-up paragraphs.
Scan mode — try it
Paste this into the Draft textarea, click Analyze draft:
The morning shift is short — three hours, but the work is not. We open the shutters, count the cash, and brief the next guard before handing over the keys.
Expected: the triadic clause “we open the shutters, count the cash, and brief the next guard” is underlined in the draft and shown as a Triadic packaging card on the right. The em-dash is not flagged — punctuation alone is not a registered pattern.
The six cases
| Case | Observed result |
|---|---|
| 1. Compare identical drafts | No added or modified findings, as expected. |
| 2. Keep one sentence; rewrite another | Retained sentence marked unchanged; rewritten sentence marked modified, as expected. |
| 3. Paraphrase without exact anchors | Correspondence remained unknown; no added or modified findings, as expected. |
| 4. Scan the morning-shift example above | One triadic finding; no em-dash finding, as expected. |
| 5. Scan a paragraph about a cat and sparrow | One triadic finding. We had expected none; that expectation was mistaken. |
| 6. Scan plain prose without a registered pattern | No findings, as expected. This control was added after case 5. |
A mistaken expectation, disclosed
Case 5 was originally designed as a “no findings on neutral prose” control. Its input was:
The cat slept on the windowsill all afternoon. A sparrow landed on the railing, looked in, and flew off.
The expectation was wrong: the second sentence contains a three-verb parallel clause, which the detector correctly flags. A structural triad in clean prose is not a product defect — it is the kind of pattern the tool is built to surface. Case 6 was added afterwards as the true negative control.
Limits
- The detector set is small. A match means the pattern is registered, not that it matters for your draft.
- The tool does not summarise, rewrite, or critique. It surfaces a matched span and the pattern’s short description.
- Compare relies on exact/stable anchors. Substantially paraphrased material is marked
UNKNOWN, not silently added or modified. - The default Compare filter is Added/modified only. Patterns that appear on both drafts don’t help you see what the second draft added or modified, so they are filtered out.
- The cases use synthetic short paragraphs. Real writing is longer and more varied; the patterns that surface will differ.
What it does not do
- It does not prove AI authorship.
- It does not tell you whether the prose is good.
- It does not generate rewrites. Noticing a pattern is itself useful, but the writer does the actual rewriting.
Status
Experimental. Built and operated by Brent.