What does this skill do, and when should you use it?
kanban-ai is an open-source (Apache-2.0) plugin for Claude Code that implements a Kanban board as a set of Markdown card files in a kanban/ directory. Each card is a .md file with status tracked in YAML frontmatter, and the board view is derived at read time. The skill supports five status columns (backlog, todo, doing, done, archive), manages dependencies via a blocked_by field, and requires appending a durable narrative record to each card's ## Narrative section. Bundled bash helper scripts handle board viewing, tag search, and blocked-card reports.
Reads and writes Markdown card files in a kanban/ directory: creates cards (with auto-incrementing ids, kebab-case filenames, and optional Job Stories); moves cards by updating the status field in frontmatter, first verifying that every card in blocked_by is done before allowing a move to doing; relocates done cards into kanban/archived/; appends ISO-dated narrative entries recording reasons, discoveries, and decisions; and runs bundled bash scripts (view_board.sh, search_by_tag.sh, search_content.sh, show_blocked.sh, list_tags.sh, list_all_cards.sh) to display the board grouped by status, search by tag or content, and list blocked cards and tag counts.
- A developer using Claude Code who wants project-local task tracking in plain-text files that live in git
- A small team that wants board state committed, branched, and reviewed alongside the code
- Users who need task dependencies: the blocked_by field prevents cards from entering doing before their prerequisites finish
- Anyone who wants a durable decision log: each card's Narrative section records why things happened, not just status changes
- Users who want to query the board directly from a terminal by tag, keyword, or blocked status without invoking an AI
How do you install this skill?
- Static review only; no scripts were executed, all runtime behavior is inferred, confidence is low.
- Cards moved to kanban/archived/ are not matched by the scripts' top-level *.md glob, so archived cards may be invisible to board/search scripts despite SKILL.md's id-scan rule including them.
- Frontmatter parsing relies on grep; values containing spaces (assignee, comma-separated tags) may parse incorrectly. Publisher identity is unverified by the registry.
- Documentation is English-only with no Chinese support; concurrency and large-board boundaries are undeclared.
- Shell / CLI
- Local filesystem
Claude Code v1.0.33+ (for plugin install)bash (for bundled helper scripts)
Requires Claude Code v1.0.33+ (check with claude --version). Two options: 1) From a marketplace that carries it: run /plugin install kanban@marketplace-name inside Claude Code; 2) From a local clone: git clone https://github.com/mattjoyce/kanban-skill.git then claude --plugin-dir ./kanban-skill (multiple --plugin-dir flags can be combined). Verify by running /help inside Claude Code and confirming kanban:kanban-ai is listed.
tmp="$(mktemp -d)"
git clone --depth 1 https://github.com/mattjoyce/kanban-skill.git "$tmp"
mkdir -p ~/.claude/skills
cp -R "$tmp/skills/kanban-ai" ~/.claude/skills/
rm -rf "$tmp"Generated from the source repository and skill path; it copies only this skill's folder. If the author's install steps above differ, follow those first. To scope it to one project, replace ~/.claude/skills with that project's .claude/skills.
How do you use this skill?
Once installed, send your agent any of these to trigger it:
- create a new kanban card for implementing user auth
- show me the board
- move card 3 to doing
- what cards are blocked?
The skill is model-invoked: just talk about tasks or the board in Claude Code, e.g. 'create a new kanban card for implementing user auth', 'show me the board', 'move card 3 to doing', 'what cards are blocked?'. Cards live in a kanban/ folder in your project root (Claude creates it if missing). You can also bypass the AI and run scripts directly: bash skills/kanban-ai/scripts/view_board.sh kanban/ — all scripts take the kanban directory as the first argument, defaulting to the current directory.
What are this skill's strengths and limitations?
- Zero infrastructure: no database or server — the board is just Markdown on your filesystem
- Board state is version-controllable with git: commit, branch, and review it like code
- blocked_by dependency checks prevent cards entering doing prematurely
- The Narrative section preserves decision rationale and outcomes for future readers
- Bundled bash scripts allow direct terminal queries without AI involvement
- Install docs cover only Claude Code (v1.0.33+); other Agent Skills clients require manually loading the SKILL.md
- No network or cloud sync — multi-device collaboration only via a shared git repo
- No locking or conflict handling for concurrent edits, an inherent limitation of plain files
- No test suite or adoption evidence (stars/downloads) visible in the source; maturity is unverified
- Fixed five-column workflow (backlog/todo/doing/done/archive); custom columns are not supported
How does this skill compare with similar options?
Side by side with related skills; every score comes from the same FSRS standard.
| Skill | FS score | Stars | Last updated | License |
|---|---|---|---|---|
| Kanban AI this page | 52 · Use with care | ★ 25 | 7mo ago | Apache-2.0 |
| Markdown to EPUB Skill | 52 · Use with care | ★ 167 | 1y ago | MIT |
| Obsidian Vault Organizer | 49 · Use with care | ★ 10k | 11d ago | MIT |
| Notion Spec-to-Delivery Planner ✓ OpenAI · Official | 43 · Not recommended | ★ 28k | 4mo ago | — |
| Obsidian CLI Assistant | 46 · Use with care | ★ 50k | 1mo ago | MIT |
The README positions it as a Kanban plugin for Claude Code; unlike hosted boards such as GitHub Projects or Jira, its differentiator is that all data stays in local Markdown files operated by an AI assistant — at the cost of hosted platforms' UI, notifications, and collaboration features.
How did FollowSkills review this skill?
The skill only reads/writes/moves markdown card files inside a local project directory; no network calls, no credential access, no covert exfiltration; archiving is a reversible file move, the narrative rule forbids deleting prior content, and Job Stories require user confirmation — data flow is transparent. Deductions: no explicit user-confirmation or rollback guidance for card moves/narrative edits, and publisher identity is unverified with limited attribution.
The six scripts match SKILL.md's descriptions, argument conventions are uniform (first arg = kanban dir, sane defaults), and view_board.sh reports a missing directory with an error and non-zero exit. Deductions: no tests or committed execution evidence; frontmatter extraction via grep can behave inconsistently with values containing spaces, multiple H1s, or grep's multi-file prefixes; archived cards under kanban/archived/ are not matched by the scripts' *.md glob, an inconsistency with the SKILL.md id-scan rule that includes archived cards.
Triggers are fairly specific in the frontmatter (create/move/view/manage cards), scenarios are clear, and the pure-markdown design has no external service dependency, so it works offline. Deductions: capability boundaries are undeclared (concurrent edits, board size), no Chinese-language support or documentation, and the semantic boundary that archive-directory cards are invisible to most scripts is not disclosed.
Structure is clean (SKILL.md plus layered scripts), README covers install, verification, field table, and rules; Apache-2.0 license is present and a changelog is referenced. Deductions: CHANGELOG content not shown, no version number on the skill itself, maintenance ownership and update path unstated, the archived-card assumption is undocumented, and there is no FAQ or known-limitations section.
The core tasks (create/move/view/search cards) have clear workflows and scriptable output formats with real marginal value over manual organization. Deductions: static review cannot verify key paths actually work, no third-party execution evidence, and output quality on edge-case inputs (tags with commas/spaces) is unproven.
Claims are cross-checkable against committed source: README's fields/rules/script list align with SKILL.md and the scripts, providing auditable primary material. Deductions: no test suite, no CI evidence, no independent reproduction; install and runtime behavior can only be inferred, and fact/inference separation is not explicit.
Open a dimension to read why it scored that way
Evidence confidence：Low — Mostly static review, author material or a limited demo; useful for discovery, not high-risk decisions.
See the full review method →