A writing editor with an AI side-channel

Your draft is
the interface.

Drop an instruction beside any passage and the agent rewrites it in place. No chat window, no copy-paste — context and control stay where you write.

The results suggest a correlation between

input diversity and model robustness, though

the mechanism is not fully understood.

[[make this argument sharper and cite the Wang et al. finding]]

Further experiments should examine whether

this effect generalizes across architectures.

Agent edit

read_doc → edit_doc

The results demonstrate a statistically significant correlation between input diversity and model robustness, consistent with Wang et al.’s (2024) finding that heterogeneous training distributions improve out-of-distribution generalization by 12–18%.

✎

Inline instructions

Write [[...]] anywhere in your draft. The agent reads the surrounding context and makes the edit right there.

📋

Track changes

Every agent edit shows as a reviewable diff. Accept, reject, or retry with feedback — just like working with a co-author.

📄

Plain markdown

Your files stay as .md on disk. No lock-in, no proprietary format. Git-friendly, portable, yours.

Not another AI chat window.

Most AI writing tools bolt a chatbot onto a text editor. You copy a paragraph out, prompt for changes, copy it back. DocWriter inverts this: your draft is the prompt. The agent reads the document, sees your instructions in context, and transacts edits through a live collaborative document — the same CRDT your cursor is on.

Get early access