Jujutsu (jj) Usage Guide for Claude Code Skill

Jujutsu (jj) Usage Guide for Claude Code

This document instructs Claude Code to use jj (Jujutsu) instead of git for version control operations. jj is a Git-compatible VCS that provides a simpler mental model and powerful history editing.

Core Concepts

Key Differences from Git

No staging area - Every change is automatically part of the working copy commit
Working copy is a commit - The @ symbol represents your current working copy commit
Anonymous branches - Commits don't need branch names; use "bookmarks" when needed for pushing
Conflicts are values - Conflicts don't block operations; they're tracked and can be resolved later
Automatic rebasing - Descendant commits automatically rebase when you modify history

Terminology Mapping

| Git | jj | | --- | --- | | branch | bookmark | | HEAD | @ (working copy) | | checkout | edit or new | | stash | Not needed (just create new commits) | | staging/index | Not applicable | | commit --amend | Just edit, changes auto-apply to @ |

Common Operations

Viewing State

# Show current status (like git status)
jj status
jj st              # short form

# Show commit log (like git log --graph)
jj log
jj log -r 'all()'  # show all commits

# Show diff of working copy
jj diff --git

# Show diff of specific commit
jj diff --git -r <commit>

# Show commit details
jj show <commit>

Creating Commits

# Describe the current working copy commit
jj describe -m "feat: add new feature"
jj desc -m "feat: add new feature"  # short form

# Create a new empty commit on top of current
jj new
jj new -m "feat: starting new work"  # with message

# Create new commit on top of specific commit
jj new <commit>
jj new main        # new commit based on main

Editing History

# Edit an existing commit (moves @ to that commit)
jj edit <commit>

# Squash current commit into parent
jj squash -m "combined message"

# Squash specific commit into its parent
jj squash -r <commit> -m "combined message"

# Squash specific paths from one commit into another (no editor)
jj squash --from <src> --into <dst> <paths>

# Restore specific paths to the state at another commit
jj restore --from <rev> <paths>

# Split a commit by file paths (no editor, see "Avoiding Interactive Editors")
jj split <path>...

# Absorb changes into appropriate ancestor commits
jj absorb

# Rebase commits
jj rebase -r <commit> -d <destination>      # single commit
jj rebase -s <commit> -d <destination>      # commit and descendants
jj rebase -b <commit> -d <destination>      # whole branch

Working with Bookmarks (Branches)

# List bookmarks
jj bookmark list

# Create a bookmark at current commit
jj bookmark create <n>
jj bookmark create <n> -r <commit>  # at specific commit

# Move a bookmark to current commit
jj bookmark set <n>

# Delete a bookmark
jj bookmark delete <n>

# Track a remote bookmark
jj bookmark track <n>@origin

Remote Operations

# Fetch from remote
jj git fetch
jj git fetch --remote origin

# Push bookmark to remote
jj git push --bookmark <n>
jj git push -b <n>  # short form

# Push current commit's bookmark
jj git push

# Create and push in one step (if bookmark exists)
jj bookmark set feature-x && jj git push -b feature-x

Handling Conflicts

# Conflicts don't block operations - check for them
jj log -r 'conflicts()'

# Resolve conflicts in working copy
# Just edit the files, remove conflict markers, save

# After resolving, the commit auto-updates
jj status  # verify resolved

Undo and Recovery

# Show operation log
jj op log

# Undo last operation
jj op undo

# Restore to specific operation
jj op restore <operation-id>

Workflow Patterns

Starting New Work

When starting substantial new work (feature, bug fix, refactor), create a new commit first:

# Start new work from main
jj new main -m "feat: description of the work"

# Or start from current position
jj new -m "fix: description of the fix"

This keeps changes isolated and makes it easier to:

Understand what changed for a specific piece of work
Review changes before merging
Revert or modify specific work without affecting other changes

Making Changes

The cheapest way to get atomic commits is to create them as you go, not to split a sprawling working copy afterward. Between logical steps, run jj new:

# start work with a description already in place
jj new -m "refactor: extract auth helper"
# ...edit files...

# move on to the next logical step
jj new -m "feat: use auth helper in login flow"
# ...edit files...

# and the next
jj new -m "test: cover auth helper edge cases"

If the description needs updating later:

jj describe -m "refactor: extract auth helper from login module"

If two adjacent commits turn out to be one logical change after all:

jj squash -m "combined message"

Cleaning Up Before Push

# Squash fixup commits into their parents
jj squash -r <fixup-commit>

# Or use absorb to automatically distribute changes
jj absorb
# Note: absorb only moves hunks into ancestor commits that already
# touch the same lines. It won't create new commits; if you have
# mixed unrelated changes in @, split them first.

# Rebase onto latest main
jj git fetch
jj rebase -d main

Creating a Pull Request

# Ensure bookmark exists
jj bookmark create my-feature

# Push to remote
jj git push -b my-feature

# Or push and create bookmark tracking
jj git push --bookmark my-feature

Updating a PR After Review

# Edit the commit that needs changes
jj edit <commit>

# Make changes (they apply directly)
# Descendants auto-rebase

# Return to working on tip
jj new <tip-commit>

# Force push updated bookmark
jj git push -b my-feature

Best Practices

Commit Messages

Use conventional commits: feat:, fix:, chore:, refactor:, docs:, test:
Describe the commit when the work is complete, not at the start
Use jj describe to update messages as work evolves

History Hygiene

Use jj squash to combine WIP commits before pushing
Use jj absorb to fold review feedback into original commits
Keep commits atomic and focused

Bookmark Naming

When pushing to GitHub, use descriptive bookmark names:

jj bookmark create fix/issue-123-auth-bug
jj bookmark create feat/user-dashboard
jj bookmark create claude/<feature>-<session-id>  # for Claude Code sessions

Before Push Checklist

# 1. Check for conflicts
jj log -r 'conflicts()'

# 2. Ensure clean status
jj status

# 3. Rebase onto latest main
jj git fetch && jj rebase -d main

# 4. Review changes
jj log -r '::@ ~ ::main'

# 5. Push
jj git push -b <bookmark>

Coexistence with Git

jj can coexist with git in the same repository:

# Initialize jj in existing git repo
jj git init --colocate

# jj and git now share the same .git directory
# You can use either tool, but prefer jj for day-to-day work

When colocated:

jj git fetch updates both jj and git refs
jj git push pushes via git
git commands still work if needed for edge cases

Avoiding Interactive Editors

jj commands like squash, describe, commit, and split open an interactive editor by default. Since LLMs cannot drive TUIs, always use the non-interactive forms:

Commit messages inline: always pass -m "message" to describe, squash, commit, new.
Squash without editor: jj squash -m "message" or jj squash --into <rev> -m "message".
Squash specific paths only: jj squash --from <src> --into <dst> <paths> moves only the named files, no editor involved.
Restore specific paths: jj restore --from <rev> <paths> pulls paths back to the state at <rev>, useful for peeling changes off a commit.

Splitting commits non-interactively

jj split can be used non-interactively — don't avoid it, just use the right invocation:

File-based split (preferred): jj split <path>... moves the listed paths into the first commit and leaves everything else in the second. No editor opens. Use this whenever atomic boundaries align with files.
Parallel split: add --parallel if the two halves should be siblings rather than stacked.
Messages inline: add -m "first" -m "second" to set both commit descriptions without prompting.

Example:

jj split -m "refactor: extract helper" -m "feat: use helper in X" \
    src/helper.rs

When file-based split isn't enough

For hunk-level splits within a single file, prefer one of these patterns over trying to drive the interactive TUI:

Split before you mess up — the best strategy. Between logical steps, run jj new -m "next thing". You end up with atomic commits by construction and never need split.

Compose with squash/restore:

jj new              # empty change on top
# edit files to the desired state of part A
jj squash --from <original> --into @ <paths-for-A>
# the original commit now contains only part B

Patch round-trip (last resort, fully scriptable):

jj diff -r <rev> --git > /tmp/full.patch
# split the patch into partial.patch and rest.patch
jj restore -r <rev>
jj new -r <rev>-
git apply /tmp/partial.patch && jj commit -m "part A"
git apply /tmp/rest.patch     && jj commit -m "part B"

Command Quick Reference

| Action | Command | | --- | --- | | Status | jj st | | Log | jj log | | Diff | jj diff --git | | Describe commit | jj desc -m "message" | | New commit | jj new | | Edit old commit | jj edit <rev> | | Squash into parent | jj squash -m <message> | | Squash paths | jj squash --from <src> --into <dst> <paths> | | Split by file | jj split <path>... | | Rebase | jj rebase -d <dest> | | Create bookmark | jj bookmark create <n> | | Fetch | jj git fetch | | Push | jj git push -b <n> | | Undo | jj op undo |

Error Recovery

If something goes wrong:

# See what happened
jj op log

# Undo the last operation
jj op undo

# Or restore to a known good state
jj op restore <operation-id>

jj's operation log means you can always recover from mistakes - nothing is ever truly lost.

Agent Skills: Jujutsu (jj) Usage Guide for Claude Code

Install this agent skill to your local

Skill Files