markwise
GitHub

The review layer for agent-written markdown.

Your agent writes the doc.
You comment and suggest edits on the exact text.
Your agent reads every note, acts, and answers.

Paste this prompt into Claude Code, Codex, or any coding agent.

Install Markwise for me with npm i -g markwise, then run markwise agent-setup and follow what it prints.

How the loop closes

Five steps, one file. The document stays put; only its state changes - from the agent's first draft to a resolved note saved inside the markdown.

  1. 1

    Your agent writes the doc.

    Ask in plain language. Claude Code, Codex, or any agent drafts launch-plan.md the way it already works.

  2. 2

    markwise opens it in your browser.

    One command turns the file into a clean reading surface. The writing leads; the tooling stays out of the way.

  3. 3

    You comment on the exact words.

    Select any span and leave a note or a suggested edit, anchored to the text it is about.

  4. 4

    Hand off kicks off the agent.

    One click bundles every open note. Your agent revises the doc and answers each one in-thread, as suggestions you can accept.

  5. 5

    You resolve. It stays in the file.

    Every comment, edit, and reply lives inside the markdown as hidden HTML comments, invisible in any normal preview.

The previewer, on a real brief

Not a mockup: this is markwise preview running on an agent-written doc, with six open notes.

The Markwise previewer: a product brief with inline suggested edits (struck text, replacements, green inserts, comment highlights) and a rail of six notes labelled replace, insert, comment, and delete, plus a Hand to agent button.

Inline suggested edits, anchored comments, a notes rail, and a "Hand to agent" button that bundles every open note. Three themes; this is Light.

markwise preview

Review in the browser: comment, reply, suggest edits, resolve.

markwise prompt

Bundles the doc and every open note into an instruction block any model can act on.

markwise status

Whose turn is it? Open vs resolved, waiting on you vs waiting on the agent.

markwise lint

24 rules guard the review data; --fix repairs mechanical drift, never prose.

markwise export

A clean copy with every trace of review stripped. The original is never touched.

✓In-file truth

The document and all its review state are one self-contained artifact. No sidecar files, no database, no platform.

✓Clean preview

Review data hides in HTML comments. GitHub, VS Code, and every normal markdown preview render the document clean.

  • Not Google Docs: your doc would leave the repo, and the agent loses it.
  • Not PR review: feedback would anchor to diffs in the forge, not the artifact.
  • Not CriticMarkup: editorial syntax, not a review loop - and it leaks into previews.
  • Not chat: describing locations instead of pointing at them.

Hand it to your agent

Install Markwise for me with npm i -g markwise, then run markwise agent-setup and follow what it prints.

Works with Claude Code, Codex, or any coding agent.

View on GitHub

What your agent will do with it:

  1. Install the CLI from GitHub (open source, MIT).
  2. Run markwise agent-setup and read the printed instructions.
  3. Teach itself the loop: write a doc, open markwise preview for you, then act on every note you leave.