1. Create a workspace
DocWriter works on any folder. For this walk-through, create a small one:2. Launch DocWriter
http://localhost:7723. The file tree on the left
shows essay.md.
3. Open the file
Clickessay.md in the file tree. The contents appear in the editor in
the center. You can edit any text directly; keystrokes save
continuously.

4. Leave a note for the agent and keep going
Anywhere in the document, type a bracketed note:[[ ... ]], handles
it, and deletes the directive text from the document. See Inline
directives.
If you’d rather not wait the three seconds, Cmd+Enter (macOS) or
Ctrl+Enter (Linux/Windows) in the editor wakes the agent
immediately. For requests that don’t anchor to a spot in the doc
(“research X,” “restructure across files”), click the Send button
in the top-right to open a chat popover.
5. Watch the edit arrive while you write
The agent runs in the background. You can keep typing while it works. Edits flow in as they happen, and your typing isn’t blocked. When the agent finishes its revision, the change appears in your document as a suggestion: the new text is highlighted, and a card appears in the outline pane on the left titled “Pending edit”.
6. Accept, reject, or retry
The card has three buttons.- Accept keeps the change.
- Reject discards it.
- Retry with feedback opens a small prompt. Type what was off about the agent’s edit (“too short”, “kept the wrong example”, “use past tense”) and submit. The agent reruns with your feedback.
- Click the card to scroll the editor to the spot the edit applies to. Useful when you have several pending edits across a long document.
- Acceptance is FIFO. If the agent proposes a sequence of changes, only the oldest pending card shows an Accept button. Accept (or reject) the oldest one and the next one becomes actionable. This avoids the ambiguity of “accept change 2 while change 1 is still pending” since change 2 might depend on change 1.
7. Inspect the history
The history pane on the right (toggle with the menu bar or the button at the top-right) shows every tool call the agent made: which files it read, what edits it proposed, and any errors. Useful when you want to understand how the agent reached its result.Where to go next
You have run through one full editing cycle. From here:- Tour the interface. What every part of the screen does.
- Steering the agent. Short prompts, model selection, autonomy levels.
- Reviewing edits. Accept and reject in detail, undo, multiple pending rounds.
- Hooks. Run shell commands on agent events (recompile a PDF, push to git).
- Writing a research-backed blog post. The marquee end-to-end walk-through.
- Editing an Overleaf LaTeX project. Same loop on a LaTeX manuscript, with PDF preview and git push back to Overleaf.