The Workshop: A Manuscript-Level Open Collaborator

The editorial bots are narrow by design — single chapter, single purpose, forbidden from creative territory. The Workshop is the deliberate inverse: the place an author goes to actually think with AI about their book. It's the first manuscript-level bot on UnboundProse, and the first open-ended one. There's no rubric, no canned workflow, no two-phase ceremony. There's a text box. The author asks anything.

The core idea: the author chooses what the AI sees

The Workshop is a three-pane page — a thread sidebar, the conversation, and a context panel. The context panel is the heart of it. It lists every chapter in the manuscript with checkboxes, a "Select all," and a metadata toggle. For each question, the author decides what goes into context. Ask about a single chapter's pacing with just that chapter selected. Then, on the next turn, add chapters 3 and 12 to ask whether a reveal lands given what was set up earlier. The selection persists between turns until the author changes it, and every turn records exactly which chapters and metadata it drew on.

That per-question control is what makes an open-ended manuscript bot tractable. A novel is too large to stuff into a model's context wholesale, and most questions only need a slice. Letting the author curate the slice keeps each call focused, affordable, and honest about what the answer is based on.

An open collaborator — on purpose

Where the editorial bots are forbidden from creative work, the Workshop is explicitly allowed to do it. The author opted in, they're driving, it's their private workspace. So the Workshop will brainstorm a plot turn, pressure-test whether a scene is working, critique dialogue, and draft or rewrite a passage on request.

This looks like it contradicts the "never touch the creative work" principle behind the editorial bots. It doesn't — and naming why was part of the design. That principle governs unsolicited, automated review: findings the author didn't ask for, generated on their prose without invitation. The Workshop is the opposite: a conversation the author starts, controls turn by turn, and can ignore. The bot still frames changes as suggestions — the author's voice stays the author's — but it doesn't withhold help the author explicitly asked for. Different mode, different rules, and the system prompt says so out loud.

Many threads, because thinking has many tracks

A single perpetual chat would pile every line of thinking into one history. The Workshop instead gives the author as many named threads as they want — one for a character's arc, one for pacing, one for a thorny chapter — each with its own title and history. It's the difference between a notebook and a single endless note.

The firewall: the AI only ever sees live, current text

The most important rule in the Workshop is invisible in the UI. Every bot reads only the live, current content of the chapters the author selects. Author-created snapshots are never sent to any model, under any circumstance.

Snapshots are a separate, human-only feature — versioned drafts an author saves and reviews side by side. Walling them off from every AI call is a deliberate trust boundary: the author can keep private drafts and version history that the AI provably never touches. Each Workshop turn stores a selection record — which chapters and metadata it used — but never a copy of a snapshot. "What did this answer draw on?" is always answerable; "did the AI read my private draft?" always has the same answer: no.

Honest about cost

The Workshop runs on the author's own Anthropic API key. A whole-book question costs materially more than a single-chapter one, so the context panel shows a live token estimate as chapters are checked, shifts to a warning when the selection gets large, and disables sending when a selection would exceed the model's window — telling the author to trim rather than silently truncating. No hidden costs, no surprise truncation, no pretending the model saw more than it did.

The pattern underneath

The Workshop is a study in giving users control of the context window instead of hiding it. Most AI chat surfaces decide for you what the model can see. UnboundProse hands that decision to the author, makes the cost of it visible, and draws a hard line around the data the model is never allowed to read. The result is an open-ended creative tool that a careful writer can actually trust.