Author: send a paragraph, get a finished draft

The draft that isn't yours

So you've pasted your notes into an AI tool and asked for a draft. What comes back is grammatical, well-structured, and confident, and you still can't send it. It doesn't sound like you: the words are the ones every model reaches for, the rhythm is the same rhythm as everyone else's announcement, and anyone who reads your writing regularly will notice the switch. Editing it back into your own register takes about as long as writing from scratch, so after a week or two you quietly go back to writing from scratch.

For a company, this is more than an annoyance. Posts, announcements, and marketing are the writing your customers judge you by, and readers have gotten good at smelling template prose. A draft in the wrong voice is unusable, because voice was the hard part all along. The shaping work that AI genuinely removes, the outlining and formatting and restructuring, only pays off if the result is something you'd put your name on.

Author, a prototype from one of our design sprints, takes voice as the central problem, and it makes two commitments. Your voice is an asset the software holds: it already exists in everything you've published, and the design treats it as a named profile you can open and read, so you stop retyping it into every prompt as a plea. And the agent's authority over a draft is never larger than the scope you selected. First we show the split of the work, then the edit paths that meter the agent's reach, then the receipt: we revised this post with the same loop.

You bring the beats; a style guide carries the voice

The workflow splits the writing into two inputs. The first is the spine of the piece, which only you can supply: an outline, a TL;DR, the three points the post needs to make, in whatever rough form they exist. A paragraph dictated into your phone qualifies. The second is a style guide the system builds from your previous writing, and this is where the voice lives: the words you reach for, the words you refuse, how you open, how much you hedge, what your audience already knows.

The prototype makes this concrete in its Draft Assistant rail. The active profile is named and inspectable: House, the research-analyst profile, described as Analytical · Measured · Evidence-led. It carries an audience description, a reach-for list (thesis, unit economics, net revenue retention), and a ban list: revolutionary, game-changer, synergy, seamless, and the exclamation mark. That's the same discipline a good editor applies by hand, held as an object the software enforces on every draft.

Structure gets the same treatment. The rail runs in one of six document modes, each with its own skeleton: an exec summary runs Context · Findings · Recommendation · Next steps; a LinkedIn post runs Hook · Story · Takeaway · Soft CTA. The same pile of notes becomes a different document in a different mode, so one set of beats can ship as a memo today and a post tomorrow.

The stronger incumbents are moving this way too. Jasper's Brand Voice builds a voice profile by analyzing writing samples or a site you point it at, and Writer's voice profiles let admins attach style guides and terminology lists, approved and banned terms included, that apply across a team's outputs. General chat assistants sit at the other end of the category: voice arrives as free-text instructions you restate per conversation. Author's difference from the profile tools is where the profile sits and what else it bounds: the voice stands in a rail beside the draft you're editing, and the same design that holds the voice also scopes each edit, which is the next section.

Every edit path scopes what the agent may change

A first draft that lands mostly right is still worthless if fixing the rest is painful, so the design spends most of its care on the editing paths. There are four, offered the moment a draft appears: reply in the composer to reshape the whole thing (“make the recommendation more decisive”), click any line to edit it in place, hover a paragraph to revise that block, or select a phrase and change only the selection.

Three grips on one draft. Each edit path scopes what the agent may change: the whole draft, one line, or one phrase.

The scopes are why this feels different from a chat tool. An agent that can silently rewrite anything is an agent you can never relax around. Here, the scope you choose is a boundary the agent respects: select one phrase and it may change that phrase and nothing else; click a line and the fix happens in place, with no round trip through the model. Conversation addresses the document; selection addresses a span. Selecting a phrase before you speak authorizes a change of exactly that size, and what makes the small scopes workable is a capacity no earlier tool had: rewriting an arbitrary span so it still agrees with the sentence around it, in a voice held constant.

This is the supervision shift our fourteen-prototype essay traces across the whole gallery, applied to drafting. A word processor makes you the writer: one cursor, every keystroke yours. A chat assistant hands the whole document to the model and leaves you a message box. Author puts the agent in the chair and hands you a vocabulary of delegation (the whole draft, a block, a line, a phrase), so you supervise a draft you didn't type, and it stays yours anyway.

We ran this loop on the post you're reading

One fact worth stating plainly: this article was revised with exactly the workflow Author wraps in a product surface. I supplied the beats as a dictated TL;DR of the point I wanted each section to make. A style guide mined from my older, hand-written Medium posts set the voice rules, including the ban lists. An agent did the shaping, and the edits after that were scoped, the way the paths above describe. The prototype is a design exploration and nothing here ships today, but the split of work it proposes is the one we already write with.

The composer is the page

All of this meets you in one place on screen. There's no blank document waiting to be filled. The plain box reading “Paste or type your raw notes…” is the front door of the whole product, and the hero states the contract flatly: “Send a paragraph. Get a clean exec summary.” Sending input is as low-stakes as firing off a message, and what comes back is a structured document with your voice rules already applied. A small TEXT / VOICE toggle on the composer means the notes can be spoken instead of pasted, which matters more than it looks: the rough dictated version of your thinking is exactly the raw material this workflow wants.

The design stays small on purpose. There's no formatting toolbar, no template gallery, no attempt to be where documents live forever; a draft ends at a Copy button and a word count. Shaping notes into a draft you'd actually send is the whole job, and everything on the screen either serves it or was left out.

Built for our own posts first

We're designing Author for a job we have every week: writing AI Hero's posts and social content in a voice that's recognizably ours, from notes that start as a dictated paragraph. The same design customizes per customer. A marketing team gets a voice profile mined from their published writing, their document modes, their ban list; a firm whose newsletters should sound like its founder holds that voice as a durable object instead of a plea retyped into every prompt.

The other consequence is what the conversation is for. With the voice standing in the profile and the scopes holding each edit, the chat is reserved for what's genuinely new about this piece, the way you'd brief a writer you trust. If you want to feel the loop yourself, open the prototype above and send it a paragraph of your own notes.