All posts

Craft

Open Atelier is open

A writing teacher that reads your book, marks what holds a chapter back, and teaches the fix on your own pages. What it does, what it will not do, and what I got wrong on the way.

By Jean-Philippe LeBlanc4 min read

The first commit went in on November 8. Six days and a little over three hundred commits later, Open Atelier is live at openatelier.co, and this weekend it goes up on Product Hunt.

It is the next version of AIStoryHub. The old one did a lot of things. This one is built around one loop, which I wrote at the top of the project brief before any code existed: diagnose, learn, practice, re-check.

What that means on a page

You paste a chapter, or write one in the editor. The check reads it and marks what holds it back. Some marks are about dialogue: a run of tags like "she hissed" and "he growled" where "said" would have disappeared, or a speech that goes on for a page without a breath. Some are about point of view, like "she felt afraid" where the reader could have watched her be afraid. Some are about continuity, like a character called Mara in chapter two and Marra in chapter nine.

Each mark opens a short lesson on that exact problem, beside the chapter, not in another tab. Then you practice on your own passage. The rest of the chapter dims, and the rules read the passage again each time you pause. When you are done, you check again and see what changed.

That last step is the one I care about most. Most writing tools stop at the diagnosis. A red mark tells you something is wrong. It does not tell you whether your fix worked, or whether you swapped one showy tag for another (the comparison counts that as still there, on purpose).

There are fourteen lessons so far, across dialogue, point of view, prose, scene, character and continuity. I wrote their examples from scratch. None of them quote published fiction.

On AI

I built this for writers who use AI and for writers who would rather not touch it, and I did not want to argue with either. So each manuscript has a mode, and the server enforces it.

In Human only, no model is ever called. The rule-based checks still run, and they are most of what the product does.

In Coach, a model (Claude Sonnet 5.5, from Anthropic) reads the chapter and leaves notes, the way an editor would in the margin. It never writes into your text. If a note puts words in quotation marks that your chapter does not contain, the note is thrown away, because at that point it is rewriting, and Coach does not rewrite.

In Co-write, it can suggest passages. Nothing goes in until you accept it, and what you accept is marked as AI-generated for good. Editing it later does not wash the mark off.

Every span of text keeps a record of how it arrived: typed, pasted, imported, or from the model. I want to be careful about what that record is. It is a record, not proof. Anyone can retype a paragraph. It exists so you can be truthful about how you wrote, if you want to be, and nobody can use it to certify you.

Your pages are never used to train a model. Model requests go out with zero data retention.

What I got wrong

A few things, in a week.

I built share cards, a little image showing your findings drop from twelve to four. Then I built links that let an editor mark each finding agree or disagree. I removed both the same day. They were about the product, not about the book. What replaced them is plainer: a private link that lets someone read your manuscript, with a password if you want one, and, if you ask, tell you where they got lost.

The rules were too eager at first. A check that flags every "just" and every "suddenly" in a published novel is noise, so I ran the rules against public-domain novels and raised the thresholds until professional prose drew few findings. They will still be wrong sometimes. You can keep any finding on purpose, with a reason, and it stops nagging you.

And I priced it wrong. Ask the book, where you ask a question about your own manuscript and get an answer that links to the paragraphs it rests on, could cost more in model calls than a subscriber paid. A margin review caught it, and I repriced. Atelier is now $21 a month.

Trying it

The check works on the home page without an account. Paste a page and see what it marks. The editor, the rule-based checks and the lessons are free with no limit, plus a few Coach reads a month. If you come from Product Hunt (openatelier.co/?ref=producthunt) and sign up before the end of November 22, Pacific time, you get 20 more Coach reads.

What I want most from this launch is to hear which findings felt right and which felt wrong. A finding that makes you roll your eyes is a bug. Write to support@openatelier.co, or open a request from inside the app. They come to me.

Check a page of your own

Paste a page and see what the check finds. It runs in your browser, with no account.

Try it on your chapter
Runs in your browser. Nothing you paste is sent. No AI.Tuned on published novels, so good prose draws few flags.