Working with Claude Code¶
Claude Code gets a first-class, auto-installed integration. Running kanbai init
writes everything Claude needs into .claude/:
.claude/rules/kanbai.md— teaches Claude the board convention and workflow. Loaded every session, so Claude always knows how to drive the board.- Skills — installed under
.claude/skills/:kanbai-next— work one card and stop.kanbai-sprint— work the whole sprint back-to-back.kanbai-status— summarize the current board.
The one-card-at-a-time loop¶
By default Claude works a single card and then stops for your review:
kanbai next --json— pick the next actionable sprint task (respecting order and dependencies). If it returnsnull, the sprint is empty or blocked — Claude tells you and stops.kanbai start <id>— move the card intodoing.- Implement it, reading
kanbai show <id> --jsonfor the description and acceptance criteria. If the scope or a design decision is unclear, Claude stops and asks first. kanbai review <id>— move the card intoreviewand stop. Claude never runskanbai done, and never moves a card toreviewwith failing lints, type checks, or tests — a card inreviewis a claim that the work is verified.- You review the work and approve it with
kanbai done <id>. If you instead ask for a change, Claude moves the card back todoingfirst — it never edits a card's work while it sits inreview.
This keeps you in the loop: nothing is marked complete without your approval, and you can review or commit between cards.
Work the whole sprint
When you explicitly want Claude to work several cards back-to-back, invoke the
kanbai-sprint skill (or just ask it to “work the whole sprint”). It runs the same loop
per card, still finishing each into review for your approval.
Allowing the CLI without prompts¶
Claude Code asks before running shell commands. Pre-approve KanbAI by adding Bash(kanbai *)
to the permissions.allow list in .claude/settings.json:
kanbai init prints this reminder. Editing settings.json is left to you, since it's a
sensitive file.
Planning work for Claude¶
You decide what Claude works on by curating the sprint:
kanbai add "Add OAuth login" -d "Support Google sign-in" -p high -t feature # capture in the backlog
kanbai move 004 todo # plan it into the sprint
Claude always picks a type when creating a card, evaluating which of the board's configured types fits best rather than leaving it unset.
Only cards in todo are picked up by kanbai next, so the backlog is a safe place to stash
future ideas without distracting the agent. You can also plan and reorder visually in the
web UI.
Using a different assistant? See Other AI harnesses — the board is harness-agnostic.