Skip to content
/

Tsumiki

tsumiki · classmethod/tsumiki · ★ 965 · last commit 2026-05-19

Japanese-developed Claude Code plugin delivering 40+ commands covering requirements (EARS) → design → TDD → implementation → verification for full project lifecycle.

Best whenExplicit haiku sub-agents for code exploration minimizes cost while primary model handles orchestration — first-class cost tiering per sub-agent role.
Skip ifSkipping requirements phase before implementation, Using cp instead of mv for task transitions
vs seeds
bmad-method(full lifecycle, skills-based) but Tsumiki ships both slash commands AND skills, uses Kiro-originated EARS notation for …
Primitive shape 59 total
Commands 42 Skills 14 Subagents 3
00

Summary

Tsumiki — Summary

Tsumiki is a Claude Code plugin from Classmethod (a Japanese cloud consultancy) that delivers an extensive suite of 40+ commands and 14 skills covering a full development lifecycle: Kairo (requirements → design → tasks → implementation loop), TDD (red/green/refactor), Dev Skills (context → plan → impl → verify → debug), DCS (code analysis), reverse engineering, and utility commands — all in a single installable plugin.

Problem it solves: Developers need a structured, repeatable development process that enforces requirements documentation (EARS notation), TDD discipline, and systematic code analysis without having to manually orchestrate each step.

Distinctive trait: EARS (Easy Approach to Requirements Syntax) notation enforcement in requirements phase, combined with interactive AskUserQuestion tool usage for scope selection — making requirements gathering a structured dialogue rather than a free-form prompt.

Target audience: Japanese-speaking developers and teams using Claude Code who want a comprehensive workflow covering requirements through deployment, with explicit TDD enforcement and reverse-engineering capabilities for legacy codebases.

Production-readiness: Active (965 GitHub stars, 15 contributors, v1.4.1, last pushed May 2026, MIT license).

Differs from seeds: Most similar to BMAD-METHOD (Archetype 1 — skills-only behavioral framework with full dev lifecycle) but delivered as a Claude Code plugin with slash commands (not just skills), and with explicit EARS notation for requirements (borrowed from Kiro) rather than BMAD's persona-based planning model.

01

Overview

Tsumiki — Overview

Origin

Classmethod, Inc. (a major Japanese AWS partner and cloud consulting company), released Tsumiki as an open-source MIT-licensed Claude Code plugin. Author: Makoto Kuroeda. The framework targets Japanese-speaking development teams but ships with English-accessible structure.

Philosophy

Tsumiki (Japanese: "building blocks") embodies the philosophy that development should proceed in explicit, validated building blocks — requirements → design → tasks → implementation — with each block producing a verifiable artifact before the next begins.

The framework draws heavily from Kiro's spec-driven approach (EARS notation, docs/spec/ directory structure) while adding TDD enforcement, code analysis agents (DCS), and reverse engineering capabilities for brownfield projects.

Design Choices

  • EARS (Easy Approach to Requirements Syntax) notation for requirements — structured, machine-parseable acceptance criteria
  • AskUserQuestion tool for scope selection at each requirements/design step (interactive, not assumption-based)
  • Parallel Explore sub-agents for code analysis (multiple agents in parallel for different aspects)
  • Context file generation (docs/dev/context.md) as a "Memory Bank" — a compact project state shared across all dev-* skills
  • kairo-loop command for continuous implementation with context compaction support
  • orchestrate command for automatic task decomposition and multi-agent execution
  • Cross-tool support via rulesync for non-Claude-Code environments

Manifesto-style quotes (verbatim from README)

"基本的にClaude Codeをサポートしますが、それ以外のツールでも使用できます。" (Basically supports Claude Code, but can be used with other tools too.)

"要件定義から実装まで、AIを活用した効率的な開発プロセスを提供します。" (Provides an efficient development process utilizing AI from requirements definition to implementation.)

02

Architecture

Tsumiki — Architecture

Distribution

Claude Code plugin (installable via /plugin marketplace add). Also supports non-Claude-Code tools via rulesync.

Install

/plugin marketplace add https://github.com/classmethod/tsumiki.git
/plugin install tsumiki@tsumiki
# Commands run with /tsumiki: prefix, e.g. /tsumiki:kairo-requirements

Plugin Manifest

{
  "name": "tsumiki",
  "version": "1.4.1",
  "commands": "./commands/",
  "skills": "./skills/"
}

Directory Tree

tsumiki/
├── .claude-plugin/
│   ├── plugin.json
│   └── marketplace.json
├── .claude/           # Claude Code native hooks/settings
├── commands/          # 40+ command .md files
│   ├── kairo-requirements.md
│   ├── kairo-design.md
│   ├── kairo-tasks.md
│   ├── kairo-loop.md
│   ├── tdd-requirements.md
│   ├── tdd-red.md
│   ├── tdd-green.md
│   ├── tdd-refactor.md
│   ├── tdd-verify-complete.md
│   ├── auto-debug.md
│   ├── orchestrate.md
│   ├── dcs/            # DCS sub-commands
│   └── ...
├── skills/            # 14 skill directories
│   ├── dev-context/   # Memory Bank generation
│   ├── dev-plan/
│   ├── dev-impl/
│   ├── dev-verify/
│   ├── dev-debug/
│   ├── kairo-implement/
│   └── ...
├── CLAUDE.md
├── README.md
├── DCS_README.md
└── DEV_README.md

Required Runtime

Claude Code (primary) or any Claude-compatible tool via rulesync.

Target AI Tools

  • Primary: Claude Code
  • Secondary: Gemini CLI, Roo (via rulesync export)

Config Files

  • .claude-plugin/plugin.json — plugin manifest
  • CLAUDE.md — project context for Claude Code
03

Components

Tsumiki — Components

Commands (40+ total)

Kairo Commands (full development flow)

Command Purpose
init-tech-stack Technology stack identification
kairo-requirements Requirements definition with EARS notation
kairo-design Design document generation
kairo-tasks Task decomposition
kairo-implement Implementation (uses TDD/DIRECT internally as skill)
kairo-loop Continuous auto-implementation with compact support

TDD Commands (individual TDD cycle)

Command Purpose
tdd-requirements TDD requirements definition
tdd-testcases Test case creation
tdd-red Write failing tests (Red phase)
tdd-green Minimal implementation (Green phase)
tdd-refactor Refactoring phase
tdd-verify-complete TDD completion verification

DCS Commands (analysis)

Command Purpose
dcs:feature-rubber-duck Idea organization + PRD creation
dcs:sequence-diagram-analysis Sequence diagram generation
dcs:state-transition-analysis State transition analysis
dcs:impact-analysis Change impact analysis
dcs:incremental-dev Incremental development planning
dcs:bug-analysis Bug root cause analysis
dcs:performance-analysis Performance issue investigation
dcs:code-question Source code Q&A
dcs:edgecase-analysis Edge case and error path analysis

Utility Commands

Command Purpose
help Command list and contextual help
orchestrate Auto-decompose complex tasks and dispatch to agent team
refine-plan Small-scale modification planning
refine-execute Execute refine-plan result
auto-debug Automatic test error debugging
build-fix Automatic build error fixing
env-fix Environment issue auto-fix
flaky-fix Flaky test stabilization
timeout-fix Timeout issue resolution

Reverse Engineering Commands

Command Purpose
rev-tasks Analyze existing code → task structure
rev-design Reverse-generate design document
rev-specs Reverse-generate test specifications
rev-requirements Reverse-generate requirements document

Skills (14 total, used autonomously by commands)

Skill Purpose
dev-context Auto-analyze project → generate Memory Bank (docs/dev/context.md)
dev-plan Requirements → task decomposition + implementation plan
dev-impl Test-first implementation (normal/quick modes)
dev-run Continuous auto-implementation for task range
dev-verify Batch test + build + lint verification per plan
dev-debug Category-based error diagnosis + fix
dev-navigate Navigate from goal to optimal skill
dev-init New project tech stack selection + initialization
dev-screen-spec Auto-generate/update screen specifications from source
dev-webtest-plan Generate Playwright web test plan
dev-webtest Execute Playwright web tests
kairo-implement TDD/DIRECT implementation for Kairo flow
ipa-security-check Security vulnerability check
ipa-security-guide Security best practices guide
05

Prompts

Tsumiki — Prompts

Excerpt 1: kairo-requirements.md — EARS Requirements Command (verbatim excerpt)

File: commands/kairo-requirements.md Technique: Declarative step-execution pattern with AskUserQuestion tool for scope selection, sub-agent dispatch for parallel analysis, and structured context block for variable management.

---
description: ユーザから提供された要件の概要を分析し、EARS(Easy Approach to Requirements Syntax)記法を使用して
詳細な受け入れ基準を含む要件定義書を作成します。
allowed-tools: Read, Glob, Grep, Task, Write, Edit, TodoWrite, WebFetch, AskUserQuestion
argument-hint: "[要件名] [PRDファイルパス(optional)]"
---

# context

出力ディレクトリ="docs/spec"
要件名={{requirement_name}}
PRDファイル={{prd_file_path}}
作業規模={{work_scope}}
信頼性評価=[]

# step

- $ARGUMENTS がない場合、「引数に要件名を指定してください(例: ユーザー認証システム)。PRDファイルがあればパスも指定してください」と言って終了する
...

## step2: 作業規模の確認

- AskUserQuestion ツールを使って作業規模を質問する:
  - question: "この要件の作業規模について教えてください"
  - header: "作業規模"
  - multiSelect: false
  - options:
    - label: "フル機能開発(推奨)"
      description: "詳細なEARS要件定義、包括的なユーザーストーリー、完全な受け入れ基準、非機能要件・エッジケース含む"
    - label: "軽量開発"
      description: "必要最小限の要件定義、基本的なユーザーストーリーのみ"
    - label: "カスタム"
      description: "含めたい項目を個別に選択"

Excerpt 2: dev-context skill — Parallel Explore Sub-agents (verbatim excerpt)

File: skills/dev-context/SKILL.md Technique: Parallel sub-agent fan-out for code analysis. Four Explore sub-agents run concurrently (haiku model) to analyze different project aspects simultaneously, then results are merged.

### Step 1: 並列探索

Explore サブエージェントを複数並列で起動し、プロジェクト情報を収集する。以下の4領域を並列に探索する:

#### 1a. プロジェクトルート分析
- パッケージ定義ファイルの検出: `package.json`, `Cargo.toml`, `go.mod`, `pyproject.toml`, `build.gradle`, `pom.xml` 等
- 言語・フレームワーク・ランタイムの特定

#### 1b. テスト環境分析
- テストフレームワークの検出: Jest, Vitest, pytest, go test, cargo test, JUnit 等

#### 1c. ディレクトリ構造分析
- 主要ディレクトリの役割マッピング

#### 1d. コード規約サンプリング
- 代表的なソースファイルを3-5個読み込み
- 命名規則の推定

**サブエージェント設定**: `subagent_type: Explore`, `model: haiku`(高速・低コスト)

### Step 2: 情報統合

各ファイルの読み込みには行数上限を設け、コンテキスト消費を制限する:
- CLAUDE.md の読み込み(上限: **100行**)
- README.md の読み込み(上限: **100行**)
- 設定ファイルの確認(各上限: **50行**)
09

Uniqueness

Tsumiki — Uniqueness

Differs from Seeds

Most similar to BMAD-METHOD (Archetype 1 — skills-only behavioral framework with full development lifecycle) but Tsumiki ships both commands AND skills (where BMAD is skills-only), uses EARS notation for requirements (borrowed from Kiro, not BMAD), and includes reverse engineering capabilities and explicit cost-optimized sub-agent routing (haiku for exploration). Compared to kiro, Tsumiki is a Claude Code plugin not a standalone IDE, and uses EARS notation explicitly (Kiro invented EARS-in-AI-dev; Tsumiki adopted it). Compared to spec-driver (also a skills-based framework), Tsumiki has a far larger command surface (40+ vs 24 skills) and adds DCS analysis, reverse engineering, and the Memory Bank concept. The Japanese-language primary documentation is unusual in this corpus.

Positioning

Tsumiki positions as "building blocks" for AI-driven development — a composable toolkit where each command is an independent block that can be used standalone or as part of the Kairo full flow. This distinguishes it from monolithic frameworks that require the entire workflow.

Distinctive Opinion

Explicit model tiering in sub-agentsmodel: haiku for exploration sub-agents vs primary model for orchestration — is unusual. Most frameworks use a single model for all tasks; Tsumiki explicitly designs for cost optimization at the sub-agent level.

Observable Failure Modes

  1. Japanese documentation barrier: While the framework works with English projects, commands and prompts are written in Japanese, creating friction for non-Japanese-speaking developers
  2. Token overhead: 40+ commands loaded as a plugin means significant context overhead
  3. EARS adoption curve: EARS notation is unfamiliar to most developers; the requirements phase has a learning curve
  4. rulesync dependency for cross-tool use: Non-Claude-Code use requires an additional dependency
04

Workflow

Tsumiki — Workflow

Kairo Flow (primary, requirements → implementation)

Step Command Artifact
1. Init /tsumiki:init-tech-stack Tech stack documented
2. Requirements /tsumiki:kairo-requirements docs/spec/{name}/requirements.md (EARS notation)
3. Design /tsumiki:kairo-design docs/spec/{name}/design.md
4. Tasks /tsumiki:kairo-tasks docs/spec/{name}/tasks.md
5. Implement /tsumiki:kairo-loop Source code, tests

Dev Skills Flow (existing projects)

Step Command Artifact
1. Context /tsumiki:dev-context docs/dev/context.md (Memory Bank)
2. Plan /tsumiki:dev-plan auth "feature" Task decomposition in docs/dev/
3. Implement /tsumiki:dev-run auth 001 005 Source code, tests
4. Verify /tsumiki:dev-verify auth Test/build/lint results

TDD Flow (individual cycle)

Step Command Artifact
1. Requirements /tsumiki:tdd-requirements Requirements in EARS
2. Test cases /tsumiki:tdd-testcases Test case list
3. Red /tsumiki:tdd-red Failing tests
4. Green /tsumiki:tdd-green Minimal implementation
5. Refactor /tsumiki:tdd-refactor Refactored code
6. Verify /tsumiki:tdd-verify-complete All tests passing

Approval Gates

  1. Scope selection: kairo-requirements uses AskUserQuestion to ask: Full / Lightweight / Custom feature set
  2. Code analysis confirmation: Before deep codebase analysis, asks "Is detailed code analysis required?"
  3. Task range selection: dev-run requires explicit task range (e.g., 001 005) — human specifies scope

Artifacts Per Phase

Phase Artifact
Requirements docs/spec/{name}/requirements.md
Design docs/spec/{name}/design.md
Tasks docs/spec/{name}/tasks.md or docs/dev/ task files
Memory Bank docs/dev/context.md
Implementation Source files + tests
Reverse Engineering docs/ generated documents
06

Memory Context

Tsumiki — Memory & Context

Memory Bank (docs/dev/context.md)

The dev-context skill generates a compact "Memory Bank" — a 500-line max context file containing:

  • Tech stack, test framework, coding conventions, architecture patterns
  • Build/test/lint commands
  • Key directory structure
  • Coding style examples

This file is read at the start of each dev-plan, dev-impl, dev-verify, and dev-debug skill execution, providing shared context without requiring full codebase re-analysis.

Spec Files (docs/spec/)

Kairo flow produces persistent specification artifacts:

  • docs/spec/{requirement_name}/requirements.md — EARS notation requirements
  • docs/spec/{requirement_name}/design.md — Design document
  • docs/spec/{requirement_name}/tasks.md — Task breakdown
  • docs/spec/{requirement_name}/note.md — Task note (tech stack, rules, related impl)

Cross-Session Handoff

Yes — all spec files are persisted in docs/ directory. Subsequent sessions load the spec files for context.

Context Compaction

Yes — kairo-loop explicitly supports compact mode. The dev-context skill also has per-file line limits (100 for CLAUDE.md, 50 for config files) to minimize context consumption.

Sub-agent Context Isolation

When dev-context launches parallel Explore sub-agents, each gets an isolated context window, and results are merged by the orchestrating agent. This provides parallel exploration without context window pollution.

Memory Persistence Level

Project-scoped. All memory lives in docs/ directory (git-tracked).

07

Orchestration

Tsumiki — Orchestration

Multi-Agent

Yes. Multiple sub-agent types:

  1. Explore sub-agents (model: haiku) — parallel code analysis in dev-context
  2. Task tool sub-agents — dispatched by orchestrate command and kairo-loop
  3. General-purpose sub-agents — used by kairo-requirements for task note generation

Orchestration Pattern

Hierarchical + parallel fan-out. dev-context uses parallel Explore sub-agents (fan-out), then merges results. orchestrate command decomposes complex tasks and dispatches to an agent team. kairo-loop runs sequential auto-implementation.

Execution Mode

Interactive-loop (most commands). Continuous with kairo-loop and dev-run for task ranges.

Isolation Mechanism

None (edits in-place). Sub-agents run within the same Claude Code session with isolated context windows.

Multi-Model

Yes (within Claude Code). dev-context explicitly specifies model: haiku for Explore sub-agents (fast/cheap) while the orchestrating command uses the session's primary model. This is explicit cost optimization — not just provider diversity.

Spawn Mechanism

Claude Code Task tool (subagent_type: Explore / general-purpose).

Consensus Mechanism

None.

Prompt Chaining

Yes. kairo-requirements output (requirements.md) → kairo-design input → kairo-tasks input → kairo-implement input. Each phase's artifact is the next phase's input.

Crash Recovery

Partial. The kairo-loop compact support means if context fills, the loop can restart from a known task position.

Context Compaction

Yes — explicit design goal. dev-context has line limits per file, kairo-loop has compact mode, sub-agents use haiku model to reduce cost.

08

Ui Cli Surface

Tsumiki — UI / CLI Surface

CLI Binary

None dedicated. All commands are Claude Code slash commands with /tsumiki: prefix.

Local UI

None.

IDE Integration

Primary: Claude Code. Secondary: Gemini CLI, Roo (via rulesync export).

Command Surface

40+ slash commands via /tsumiki: prefix in Claude Code. All commands defined as .md files in commands/.

rulesync Integration

For non-Claude-Code environments:

npx -y rulesync init
npx -y rulesync config --init
npx -y rulesync import --targets claudecode --features commands,subagents
# For Gemini CLI:
npx -y rulesync import --targets geminicli --features commands,subagents

Observability

  • docs/spec/ — persistent requirements/design/task artifacts
  • docs/dev/context.md — generated Memory Bank
  • Task note files per requirement (note.md)

Cross-Tool Portability

Medium. Primary Claude Code, but rulesync enables Gemini CLI and Roo support. Commands are markdown files, so theoretically any tool that reads them can use them.

Related frameworks

same archetype · same primary tool · same memory type

Claude-Flow / Ruflo ★ 55k

Eliminates single-agent context limits and sequential bottlenecks by orchestrating fault-tolerant swarms of specialized AI agents…

Hermes Agent (NousResearch) ★ 168k

Self-improving personal AI agent with closed learning loop, 7 terminal backends, and messaging gateway — not tied to any AI…

OpenCode ★ 165k

Terminal-first AI coding agent with multi-model routing, native desktop app, and a typed .opencode/ configuration system for…

OpenHands ★ 75k

Open-source AI software development platform (open-source Devin alternative) with Docker sandbox isolation, 77.6% SWE-bench…

DeerFlow ★ 70k

Long-horizon superagent that researches, codes, and creates by orchestrating parallel sub-agents with isolated contexts in Docker…

oh-my-openagent (omo) ★ 60k

Multi-provider AI agent orchestration for OpenCode: escape vendor lock-in by routing Sisyphus (Claude/Kimi/GLM) and Hephaestus…