Jujutsu (jj) is a Git-compatible version control system with automatic snapshots, undoable operations, and safe history rewriting. This practical guide shows Git users how to experiment fearlessly, recover from anything, and turn AI-generated changes into clean, reviewable history.
For years, Git was the unquestioned foundation of my workflow. But when I started
working daily with AI coding agents, the rhythm changed. Agents could generate,
rewrite, and discard large amounts of code in minutes—while more and more of my
time went into managing commits, recovering work through the reflog, and
untangling history.
Then I found Jujutsu, a Git-compatible VCS built on a different model. Nothing was ever lost, nothing needed staging, and rewriting history stopped being something I braced myself for. For the first time in years, version control felt calm.
But learning individual commands was only half the story. I looked for a guide that showed how these ideas fit together into a real development workflow with AI agents. It didn’t exist yet—so I wrote it.
reflog archaeology requiredJuju-chu! unfolds as a conversation between two engineers: a senior who knows Jujutsu inside out, and a junior who keeps asking why.
The moment a question forms in your head—“but if you don’t control when commits happen, doesn’t your history end up chaotic?” “wait, if I rewrite past history, don’t I need to rebase?”—the junior asks it on the page. Misconceptions get corrected the moment they arise, and edge cases enter the discussion as objections, not footnotes.
It’s one of the oldest teaching devices there is—Socrates used it, and so does The Little Schemer. In Japan, it’s how my Riakuto! series reached 40,000 readers. Read the sample and see how it works.
This chapter explores why Git’s human-paced model becomes a burden when AI agents make large, rapid changes. Staging gets skipped, context switches pile up, and rewriting history feels risky.
Then we introduce Jujutsu’s answer: automatic snapshots, universal undo, first-class conflicts, and consistent commands. You’ll see why Jujutsu lets you experiment first and clean up later.
After installation and setup, you’ll take a small project from initialization to GitHub—working with changes, reading history, and pushing your work. You’ll also meet Jujutsu’s three logs: the revision log, the evolution log for one change, and the operation log for the repository.
Then comes the mental model that makes everything else click: what a change is, how bookmarks differ from Git branches, and why working on anonymous branches becomes the default.
This chapter shows how to steer Claude Code and Codex CLI toward jj instead of git, automate routine commands while keeping risky ones gated, and trigger formatting and linting through hooks.
Then comes a full workflow: turning an agent’s sprawling output into reviewable changes, pushing them for a PR, and running multiple agent sessions in parallel workspaces before merging their work back together.
Start with revsets and filesets—the query languages for selecting exactly the
revisions and files you mean—then meet three power-user commands: jj absorb,
which sends fixes back to the changes they belong to, jj arrange, and
jj bookmark advance.
From there, you’ll build a custom jj push for pre-push checks, reduce manual
conflict resolution with Mergiraf and
Weave, and explore UI tools including
jjui and
JJ View.
The FAQ answers the questions Git users keep asking: why there’s no pull
command, how to do the equivalent of cherry-pick, how to keep temporary files
out of your history, and whether Jujutsu is really just a wrapper around Git.
Then come the problems that stop you cold—a plain push that silently becomes a force push, a stale working copy, a change that picked up a divergent annotation, and working-copy commit wandering off after you merge a PR and fetch the result.
Free config files: make your AI agent use Jujutsu instead of Git
git reset --hard, or destroyed by a Git command the agent ran on
its own.reflog and interactive rebase with absolute confidence, and feel
your Git workflow has zero friction.
Yuka Ooka is a Japanese technical author whose React + TypeScript series Riakuto! (りあクト！) has sold more than 40,000 copies.
She began her career as a PHP and Ruby developer and product manager before going freelance in 2016. An early adopter of React in Japan, she later worked with clients as a React specialist.
Her latest book, Juju-chu!, is the first comprehensive guide to the Jujutsu version control system—and to pairing it with AI coding agents like Claude Code.
No. The book starts with installation and builds up from there—no prior familiarity assumed. Kanae, the junior engineer in the dialogue, has never touched Jujutsu either, so you learn alongside her.
That said, it’s not only for complete newcomers to Jujutsu. If you’ve tried it before and bounced off the gap between Git’s mental model and Jujutsu’s, the book spends real time closing that gap: what a change really is, why commit IDs change while change IDs stay put, and where your Git instincts will mislead you.
Yes—this is the one prerequisite. Juju-chu! assumes you’ve used Git day to day: cloning, branching, committing, pushing, and opening pull requests. Much of the book explains Jujutsu by contrasting it with Git, so those comparisons only land if Git already feels familiar.
What you don’t need is Git mastery. If reflog is a mystery to you and
interactive rebase makes you nervous, you’re exactly who this book was written
for.
No. Jujutsu works directly with an existing Git repository, so you can adopt it on an existing project without touching anyone else’s setup. What you push to GitHub is ordinary Git commits—your teammates keep using Git, review your pull requests as usual, and never need to know you switched.
That makes Jujutsu unusually low-risk to try: the decision is yours alone, and
it’s reversible. Git is still right there underneath, so you can switch back to
git commands anytime.
No. The book uses Claude Code and Codex CLI as its worked examples, but nothing about the Jujutsu side is specific to either one.
The Codex setup is built on AGENTS.md and skills—open conventions that Cursor,
OpenCode, and other agents also read. So the rules and instructions in the book
carry over with little more than a change of filename. What you’re really learning
is how to teach an agent to use Jujutsu instead of Git, and that carries over from
tool to tool.
Jujutsu v0.43.0, released in July 2026. Every command and every log output in the book comes from an actual run against that version—nothing is written from memory or reconstructed by hand.
Jujutsu is still pre-1.0 and moves quickly, so some details will inevitably shift. But the book concentrates on the model underneath: what a change is, how the working copy is committed for you, and why history can be rewritten safely. That model is the book’s real subject—not a frozen snapshot of the CLI—and it’s the part that’s hardest to pick up from release notes.
Yes, within limits. If a breaking change in Jujutsu makes the book’s existing instructions inaccurate, I’ll revise the affected sections. I’ll also fix errors and clarify anything that could be explained better.
This is a finished book, not a continuously expanding reference. Updates will focus on keeping the existing material accurate—not covering every new Jujutsu feature or substantially reorganizing the book.
Leanpub readers get these updates at no extra charge. Updated Kindle files may also be published, though whether Amazon pushes them to existing customers is out of my hands; you can usually pull the latest version yourself from Manage Your Content and Devices.
PDF and EPUB (Leanpub)—carefully typeset with embedded fonts and completely DRM-free. The PDF preserves the full page design, while the EPUB reflows for smaller screens without sacrificing the book’s typography. If you care about how a technical book looks and reads, or you want to keep the files somewhere Amazon can’t reach, this is the one.
Kindle (Amazon)—the right pick if your reading life already lives in the Kindle app. The text reflows to fit your device, though some of the original typography is inevitably lost.
Paperback (Amazon)—6 × 9 inches, 163 pages, and the only way to get Juju-chu! in print.
Readers have also asked which purchase supports me more directly. I keep a slightly larger share of a Leanpub sale than a Kindle sale. The digital editions are the same price, so buy wherever suits you—but if it’s a coin flip, that’s the tiebreaker.