MiniSpeckit:把 spec、plan、tasks 合并成单文件的 AI 编码规范工具
saiganne/MiniSpeckit
MiniSpeckit 将传统 spec-driven 开发中分离的 spec.md、plan.md、tasks.md 合并为单个结构化文档,让 AI 编码智能体只读一个文件即可保持完整上下文,避免跨文件不同步和 token 浪费。
__ __ _ _ ____ _ _ _
| \/ (_)_ __ (_) ___| _ __ ___ ___| | _(_) |_
| |\/| | | '_ \| \___ \| '_ \ / _ \/ __| |/ / | __|
| | | | | | | | |___) | |_) | __/ (__| <| | |_
|_| |_|_|_| |_|_|____/| .__/ \___|\___|_|\_\_|\__|
|_|
A lightweight, token-efficient, zero-dependency Spec-Driven Development (SDD) toolkit for AI coding agents.
Built for Claude Code, Cursor, Copilot, Antigravity, and Codex.
⚡ Why MiniSpeckit?
GitHub's Spec Kit popularized Spec-Driven Development (SDD): giving AI agents clear specifications before writing code. But in practice, existing spec frameworks suffer from heavy ceremony and token bloat:
| Feature | Heavy Spec Kits (e.g. GitHub Spec-Kit) | MiniSpeckit |
|---|---|---|
| Files per Feature | 4 separate files (spec.md, plan.md, tasks.md, checklist.md) |
1 Single Blueprint (specs/NNN-slug.md) |
| Token Context Overhead | ~4,000+ context tokens burned on boilerplate | ~800–1,200 tokens (75% context savings) |
| Cross-Platform | Relies on Unix bash (setup.sh, new-feature.sh), breaks on Windows |
100% Cross-Platform Node.js (Windows, macOS, Linux) |
| Dependencies | Requires complex toolchains or package trees | Zero npm dependencies |
| Agent Support | Often locked into Claude Code custom commands | Universal: Claude Code, Cursor, Copilot, Antigravity, Codex |
| Progress Tracking | Manual file inspection | Built-in Terminal Dashboard & Health Checks |
🚀 30-Second Quickstart
You don't need to install anything globally. Run via npx:
# 1. Initialize MiniSpeckit in your project (works instantly via GitHub or npm) npx github:saiganne/MiniSpeckit init # or: npx minispeckit init # 2. Create your first single-file blueprint npx minispeckit new "<feature-name>" # e.g. "Dark Mode" or "Search Filter" # 3. View status and progress across all specs npx minispeckit status # 4. Lint and verify spec health npx minispeckit check
📄 The Single-File Blueprint Philosophy
Traditional spec-driven development forces developers and AI agents to juggle multiple disconnected files:
spec.md(What & Why)plan.md(Technical Architecture)tasks.md(Checklist)
The problem? AI coding agents frequently desynchronize: they update tasks.md but hallucinate requirements that conflict with plan.md, wasting context tokens and developer time.
MiniSpeckit unifies everything into one structured document (specs/001-feature-name.md):
specs/001-user-authentication.md
├── 1. Specification -> Problem statement, User Story & Acceptance Criteria
├── 2. Technical Plan -> Architecture, Files to Touch & Edge Cases
├── 3. Tasks Checklist -> Interactive - [ ] items the agent completes sequentially
└── 4. Verification -> Automated test commands and manual verification
Agents read one file, hold the entire feature in context, and tick off tasks as they write code.
🛠️ CLI Commands & Options
minispec init
Scaffolds the specs/ directory, installs default templates into .minispec/templates/, and creates AGENTS.md and agent integration rules.
minispec new <feature-name> [options]
Creates a blueprint for any feature you are building (e.g. "Dark Mode", "Checkout Flow").
Creates the next sequentially numbered spec (e.g., specs/001-user-profile.md).
Templates available:
--template default(Standard full-stack feature blueprint)--template api(Backend REST/GraphQL endpoint & DB schema blueprint)--template bugfix(Lean reproduction, root cause & regression fix blueprint)
minispec new "Stripe Webhook Handler" --template api minispec new "Fix Memory Leak in Auth Cache" --template bugfix
minispec status
Displays a clean ASCII dashboard with visual progress bars tracking completion across all specifications:
ID SPEC FILE TASKS PROGRESS STATUS
--------------------------------------------------------------------------------
001 001-user-authentication.md 5/5 [████████] 100% Completed
002 002-stripe-webhook.md 3/6 [████░░░░] 50% In Progress
003 003-profile-settings.md 0/4 [░░░░░░░░] 0% Draft
--------------------------------------------------------------------------------
Summary: 1 completed | 1 in progress | 1 draft(s)
minispec check
Audits your specs for missing acceptance criteria, empty task lists, or desynchronized status indicators.
🤖 Universal Agent Integration
MiniSpeckit comes with ready-to-use configuration files for every major AI coding assistant:
- Claude Code: Ready-made slash commands in
integrations/claude/(/minispec.new,/minispec.implement,/minispec.status). - Cursor: Instant rule preset in
integrations/cursor/minispec.mdc. - GitHub Copilot: Context instructions in
integrations/copilot/copilot-instructions.md. - Antigravity / Gemini / Codex: Universal rules in
integrations/agents/AGENTS.md.
Recommended Prompting Workflow:
- Scaffold:
npx minispeckit new "Feature Name" - Refine with AI:
"Review specs/001-feature-name.md. Help me refine the Acceptance Criteria and Technical Plan before we write code."
- Implement:
"Implement specs/001-feature-name.md step-by-step. Follow Section 3, check off tasks as you finish them, and run tests."
🧪 Testing
MiniSpeckit includes an automated test suite verifying CLI commands and template generation:
npm test🤝 Contributing
Contributions, bug reports, and community presets are welcome! Please check out CONTRIBUTING.md.
📝 License
MIT License © 2026 saiganne
来源:GitHub · github.com
