DocWriter

A harness and user interface for AI-assisted writing,
from HCI researchers at UC Berkeley.

Settings
Sidebar
No passive voice Keep paragraphs under 3 sentences Use active verbs +2 more
Why the Sky Turns Blue
.docwriter
blog-post.md
blog-post.md
đŸ± AGENT á¶»zz Send · Restart
No new request, directive, or feedback to ac...

What is DocWriter?

DocWriter is an ongoing research project at UC Berkeley. It is a local editor with a built-in writing agent. You can write any files you want—markdown, plain text, LaTeX—and the agent collaborates with you, reading your text and comments asynchronously in real time.

One of our key UI ideas is to get you out of the chat when communicating with your agent. You communicate inline in the text, e.g., [[ add a citation to ... ]], or highlight text and leave comments or suggestions for the agent to address. Every edit the agent makes appears in place, and you review it inline before it takes effect.

DocWriter works with most frontier LLMs and SDKs: Claude Code, Codex, Pi, and others. It uses your existing subscriptions, so if you already have one you do not pay extra.

Research directions

What are the research problems? Writing is different from coding—it is more personal, more taste-driven, and harder to evaluate mechanically. That means we need new harnesses, interfaces, evals, and good practices built specifically for writing with AI.

Harness engineering

What does an AI writing harness need? Turns out there are many moving parts that are hard to design well: CRDT-backed documents and custom MCP tools for agent edits, skills that encode good writing practices and the human's voice, and hooks and subagents for rules and revision passes.

Human steering of long-horizon tasks

Writing a paper or book takes many rounds of editing over months. It is not obvious how to let writers steer agents in real time over long horizons, keep quality consistent as documents grow, or enforce rules across hundreds of edits
let alone how to present changes without overwhelming the writer.

Characterizing slopwords and good practices

Everyone complains about AI “slop,” but we lack good ways to characterize it and monitor it over time, especially as new models come out. While startups like Pangram Labs sell useful slop detection products, we want to build open source frameworks that researchers and writers can use to identify slopwords and track how they change over time.

Try it

npx docwriter (coming soon!)
Plain writing skill ★ 239
DocWriter GitHub (coming soon!)

Want to get involved?

We are looking for user study participants. If you are willing to write and publish a blog post, paper, or other long-form document using DocWriter and share your experience with us, we would love to hear from you.

Reach out at [email protected].

DocWriter · EPIC Data Lab, UC Berkeley