Writing · Thinking · Open source

write-on

A writing tool that works like a thought partner. It keeps every way you tried to say something within reach, learns how you write, and nudges like a trusted editor. It never writes for you.

Try write-on Try the idea on this page Read the research report

Where it started

This project exists because of Jason Fried, co-founder of 37signals. In September 2026 he posted a short video of Write_On, a writing tool he had wanted for about a decade and built over a weekend with Claude's help.1 He described it like this:

"Essentially it's 'alternative control' at the word, sentence, and paragraph level, plus a way to dim stuff back, and stash stuff near by. Not version control, but alternative control."Jason Fried, on X

Watching that video changed how I think about writing tools. It showed me that a tool can make room for the way you actually think while you write: trying three versions of a line, keeping the one you're unsure of faded on the page, setting aside a sentence you love but can't use yet. All credit for that interface idea is his. Please watch his video first; it explains the idea better than any description.

write-on uses Write_On as its base and builds on top of it. I spent the first part of this project on deep research into what academia and industry know about writing copilots and thought partners that augment writing rather than replace it, and folded what I found into the design. The short version of that research is below. The full report is a public Google Doc.

About the name. write-on is a nod to Fried's Write_On, and also to Rock On!!, the Farhan Akhtar film and its title song. Write on, rock on.

Why

Writing is thinking

Leslie Lamport put it bluntly: "If you're thinking without writing, you only think you're thinking."2 The struggle to find the right word is not a cost to be automated away. It is where you work out what you mean.

That is why a tool that writes for you also thinks for you, and it is the one thing write-on refuses to do. The goal is to make you a better and faster version of yourself, while every sentence that ships stays yours.

A ghostwriterYou type The meeting was really very productive and we made a lot of progress. It rewrites the sentence for you.
write-onIt lets you hold The meeting moved us forward as an alternative under the same sentence. It faintly marks really very, because you cut intensifiers in most of your past drafts. If you ask, it says: What was the progress? Your best paragraphs name the specific thing. You still write the sentence.

Try it

Alternative control, in thirty seconds

Version control keeps snapshots of a whole document over time. Alternative control keeps the options for one word, sentence or paragraph in the text, right now, so you can compare them while you read. Here is a tiny working version:

We shipped it .

Then we listened to what people did with it.

Overflow (stashed nearby)

  • "Small tools, used daily, beat big tools used once."
Click a wavy word or sentence to flip to its next alternative; the dots show which one you're on and nothing is ever lost. Click the faded sentence to bring it back, or click it again to dim it. Press Use to move the stashed line into the text.

In the real editor this is all on the keyboard: hover a word or sentence and press ↑/↓ to cycle, ghost with one shortcut, send a selection to Overflow with another, and change any key you like.

What it is

Three roles, in order of how often you'll meet them

No AIThe workbenchA place to play with phrasing: three versions of a sentence, a paragraph dimmed while you decide, a line stashed for later. This is Fried's core, rebuilt.
Learns youThe attentive editorIt has read what you've written and notices when you drift from your own voice, repeat a tic or bury the lede. It points; you fix.
On requestThe sparring partnerWhen you're stuck it asks a question, offers a counterargument, or shows how you opened similar paragraphs before. A lever, never the finished sentence.

The research

What the evidence says about AI and your voice

Before writing any code, research agents went through academic papers (CHI, UIST, ICLR, EMNLP, ICML and recent arXiv preprints), commercial tools, open-source projects and decades of prior art in non-linear writing. Their notes were then synthesized and checked against the sources. Four findings shaped write-on the most.

32%
How much "preserve my voice" prompts reduced voice drift when three frontier models rewrote 300 personal stories. All three still pushed every stylistic marker the same way.3
≈2×
How much more likely people were to argue an opinionated model's side. Only 10 to 30% of them noticed the influence.4
0.74 vs 0.55
Share of suggestions writers accepted when offered three choices instead of one. Most of them preferred having several.5
17–21%
How often few-shot imitation of everyday bloggers passed an authorship check, against about 95% for news and email. Personal writing is the hardest voice to fake.6

The pattern behind these numbers is consistent. When people write with a model, their text gets more alike and more neutral, they feel less ownership, and they stay just as satisfied, so they don't notice.7 One study found that the loss of diversity came from the model's own contributions, while the writers' own words kept theirs.8 The harm enters with the machine's words.

Meanwhile the market has learned to copy your voice and then aimed that skill at ghostwriting. Every commercial tool checked uses its model of your voice to generate or rewrite text. None uses it to give you feedback. That gap is where write-on lives: use what the tool learns about your writing to notice and nudge, not to draft.

The rules that came out of it

  1. Machine words enter only as options.AI output is added as one more alternative under a span, never swapped in. Your version stays active until you flip.
  2. The AI may cut or flag, not rewrite.Trims are deletions shown as ghosts. Weak sentences are marked, not replaced.
  3. Small scope, several choices.Three alternatives at the word or sentence level, on request. No always-on autocomplete, which research shows makes writing more predictable.
  4. Your style model is yours to read and veto.A plain-language style card, every trait backed by quotes from your own writing, each one editable or deletable.
  5. Show drift; don't trust anyone to notice it.A meter shows when a piece is moving away from how you usually write.
  6. Machine text is always marked.Dimming is reserved for your own ghosts; anything a model wrote is labelled loudly.
  7. Summoned, rationed, short.The AI speaks when asked, unprompted nudges are capped, and a critique is one quoted span and one question.
  8. Plain files you own.Pieces are Markdown. Your style data stays local by default and can be exported or deleted.
Read the full research report → "Writing copilots that augment, not replace: state of the art (November 2026)". Commercial tools, HCI studies, style-learning methods, the design principles and evaluation metrics, with every source linked. Public Google Doc.

Where it stands

Built versus planned

write-on grows in four phases. Phase 0 has no AI at all, on purpose: if alternatives, dimming and stashing don't make writing better on their own, no model on top will save it. Each phase moves on only after a test on real writing, not a date.

  1. Live, still growing

    Phase 0 · Workbench, no AI

    • Alternatives at the word, sentence and paragraph level, flipped in place
    • Ghost (dim) and revive; Overflow to stash text and use it later
    • Keyboard-first, with shortcuts you can rebind
    • An event log of every edit, with a timeline you can zoom, replay and jump back into
    • Plain Markdown files; deploying as a web app
  2. Planned

    Phase 1 · Mirror

    The first AI. A cold start from your past writing builds a style card with evidence quotes. Editor personas (a New Yorker editor, an NYT editor, a favourite writer, and "you, from your own history") critique a passage and offer rewrites only as labelled alternatives you choose to flip to. Write_On's own AI ideas, such as AI alternatives and the delete-only Lab, arrive here too.

  3. Planned

    Phase 2 · Editor

    Nudges grounded in your own writing ("you usually open with a concrete scene; this opens with an abstraction"), a preference model learned from the alternatives you pick and reject, and a strict budget on interruptions.

  4. Later

    Phase 3 · Workshop

    A local model scoring as you type, an optional on-device adapter trained on your writing, sync across devices, and standing tests that check you sound more like yourself after months of use, not less.

Try Phase 0 at write-on.up.railway.app. It opens with a sample piece to play with. The code and design notes are on GitHub.

References

  1. Jason Fried, Write_On post and demo video on X, 30 September 2026. Open-source readings of the video: klipstack/undertext and joeysim/write-on-replica.
  2. A line Leslie Lamport has used in many talks. Paul Graham makes the same case in "Writes and Write-Nots" (2024).
  3. "Voice Under Revision", arXiv 2604.22142, April 2026 (preprint). 2,700 rewrites by three frontier models.
  4. Jakesch et al., "Co-Writing with Opinionated Language Models Affects Users' Views", CHI 2023.
  5. "Choice Over Control: How Users Write with Large Language Models using Diegetic and Non-Diegetic Prompting", N=129, arXiv 2303.03199.
  6. Wang et al., style imitation of 400+ everyday authors, EMNLP 2025 Findings.
  7. "How LLMs Distort Our Written Language", randomized trial, N=100, arXiv 2603.18161, March 2026; and Draxler et al., "The AI Ghostwriter Effect", on ownership.
  8. Padmakumar and He, "Does Writing with Language Models Reduce Content Diversity?", ICLR 2024.