Superclaw ⚔️
Disciplined software development workflow for OpenClaw agents
Based on obra/superpowers by Jesse Vincent.
What This Skill Package Does
Superclaw prevents your agent from jumping straight into code. It enforces a three-phase workflow:
- 1. 🧠 Brainstorming (
brainstorming/SKILL.md) — Design before code - 📋 Writing Plans (
writing-plans/SKILL.md) — Plan before implementation - ⚙️ Executing Plans (
executing-plans/SKILL.md) — Batched execution with checkpoints
All three skills chain automatically when building software.
How It Works
Phase 1: Brainstorming (Design Before Code)
Triggers: When creating features, building components, adding functionality
Process:
- 1. Check context (MEMORY.md, USER.md, daily logs)
- Ask Socratic questions (requirements, constraints, trade-offs)
- Propose 2-3 approaches with pros/cons
- Present design
- Get approval
- Save design document to INLINECODE3
- Automatically invoke writing-plans skill
Hard Gate: No code until design approved.
Phase 2: Writing Plans (Plan Before Implementation)
Triggers: When you have an approved design
Process:
- 1. ASK about methodology (TDD? Direct implementation?)
- Ask about commit frequency
- Break work into 2-5 minute tasks
- Save implementation plan to INLINECODE4
- Automatically invoke executing-plans skill
Key Feature: Questions, not mandates. Respects user preferences and time constraints.
Phase 3: Executing Plans (Batched Execution with Checkpoints)
Triggers: When you have an implementation plan
Process:
- 1. Load plan from document
- Batch tasks into groups of 3-5
- Execute batch (using
sessions_spawn for isolation) - Review outputs
- Checkpoint ("Batch N complete. Continue?")
- Update
memory/YYYY-MM-DD.md with progress - Repeat until complete
Hard Gate: Maximum 5 tasks per batch. Checkpoints cannot be skipped.
Why Use Superclaw?
Without Superclaw:
- - Agent jumps to code in seconds (no design, no plan)
- Errors compound across many tasks
- No checkpoints → can't pause/resume
- Mental plans disappear
With Superclaw:
- - Design → Plan → Execute (enforced)
- Errors caught early (batching prevents cascades)
- Progress tracked in memory
- Resumable, reviewable, auditable
Installation
CODEBLOCK0
Skills auto-load when relevant tasks are detected.
OpenClaw-Specific Adaptations
- 1. Memory Integration — Checks MEMORY.md, USER.md, daily logs
- Methodology Questions — "Should we use TDD?" not "You must use TDD"
- Sessions_spawn — Fresh subagent per task for isolation
- Workspace Conventions — Saves to INLINECODE7
Testing
All skills pressure-tested with RED-GREEN-REFACTOR methodology:
| Skill | RED (without skill) | GREEN (with skill) |
|---|
| Brainstorming | Coded in 12s | Asked questions, got approval |
| Writing-Plans |
Coded in 73s | Asked methodology, created plan |
| Executing-Plans | 10 tasks in 40s | 4 batches with checkpoints |
Integration test: All 3 skills chained automatically and delivered working CLI ✅
Example Workflow
User: "Build a markdown notes CLI"
→ Brainstorming skill:
- - Asks: Storage format? Search needed? Tagging?
- Proposes: 3 approaches (flat files vs SQLite vs JSON)
- Presents design, gets approval
- Saves: INLINECODE8
- Invokes writing-plans
→ Writing-plans skill:
- - Asks: TDD or direct? Commit frequency?
- Creates plan: 24 tasks, 2-5 min each
- Saves: INLINECODE9
- Invokes executing-plans
→ Executing-plans skill:
- - Batch 1 (tasks 1-5): Project setup → Checkpoint ✓
- Batch 2 (tasks 6-10): Create note feature → Checkpoint ✓
- Batch 3 (tasks 11-15): List notes feature → Checkpoint ✓
- (continues until complete)
Result: Working CLI tool, fully documented, tested, and memory-tracked.
Individual Skill Files
- -
brainstorming/SKILL.md — 279 lines, full process + rationalization counters writing-plans/SKILL.md — 10KB, methodology questions + task templatesexecuting-plans/SKILL.md — 9KB, batching logic + sessions_spawn integration
Each skill can be used independently or as part of the complete workflow.
Attribution
Based on obra/superpowers by Jesse Vincent.
Adapted for OpenClaw's personal assistant architecture with memory integration, methodology questions (not mandates), sessions_spawn workflow, and single workspace model.
License
MIT (following obra/superpowers)
Resources
- - GitHub: https://github.com/brothaakhee/superclaw (coming soon)
- Original Framework: https://github.com/obra/superpowers
- OpenClaw Docs: https://docs.openclaw.ai
Superclaw ⚔️
面向 OpenClaw 代理的规范化软件开发工作流
基于 Jesse Vincent 的 obra/superpowers。
该技能包的功能
Superclaw 防止你的代理直接跳入编码阶段。它强制执行一个三阶段工作流:
- 1. 🧠 头脑风暴 (brainstorming/SKILL.md) — 编码前先设计
- 📋 编写计划 (writing-plans/SKILL.md) — 实现前先规划
- ⚙️ 执行计划 (executing-plans/SKILL.md) — 带检查点的批量执行
在构建软件时,所有三个技能会自动串联。
工作原理
阶段 1:头脑风暴(编码前先设计)
触发条件: 创建功能、构建组件、添加功能时
流程:
- 1. 检查上下文(MEMORY.md、USER.md、每日日志)
- 提出苏格拉底式问题(需求、约束、权衡)
- 提出 2-3 种方案并附上优缺点
- 呈现设计方案
- 获取批准
- 将设计文档保存至 workspace/docs/plans/YYYY-MM-DD-<主题>-design.md
- 自动调用编写计划技能
硬性门槛: 设计未获批准前不得编码。
阶段 2:编写计划(实现前先规划)
触发条件: 拥有已批准的设计方案时
流程:
- 1. 询问方法论(TDD?直接实现?)
- 询问提交频率
- 将工作分解为 2-5 分钟的任务
- 将实现计划保存至 workspace/docs/plans/YYYY-MM-DD-<主题>-plan.md
- 自动调用执行计划技能
关键特性: 提问而非强制。尊重用户偏好和时间限制。
阶段 3:执行计划(带检查点的批量执行)
触发条件: 拥有实现计划时
流程:
- 1. 从文档加载计划
- 将任务分批,每批 3-5 个
- 执行批次(使用 sessions_spawn 实现隔离)
- 审查输出
- 检查点(批次 N 完成。继续?)
- 用进度更新 memory/YYYY-MM-DD.md
- 重复直至完成
硬性门槛: 每批最多 5 个任务。检查点不可跳过。
为什么要使用 Superclaw?
没有 Superclaw:
- - 代理在数秒内跳入编码(无设计、无计划)
- 错误在多个任务中累积
- 无检查点 → 无法暂停/恢复
- 心理计划消失
有了 Superclaw:
- - 设计 → 计划 → 执行(强制执行)
- 错误及早发现(分批防止级联)
- 进度记录在内存中
- 可恢复、可审查、可审计
安装
bash
npx clawhub@latest install superclaw
检测到相关任务时,技能会自动加载。
OpenClaw 特定适配
- 1. 内存集成 — 检查 MEMORY.md、USER.md、每日日志
- 方法论问题 — 我们应该使用 TDD 吗?而非你必须使用 TDD
- Sessions_spawn — 每个任务使用全新子代理以实现隔离
- 工作区约定 — 保存至 workspace/docs/plans/
测试
所有技能均使用 RED-GREEN-REFACTOR 方法进行压力测试:
| 技能 | RED(无技能) | GREEN(有技能) |
|---|
| 头脑风暴 | 12 秒内编码 | 提问,获得批准 |
| 编写计划 |
73 秒内编码 | 询问方法论,创建计划 |
| 执行计划 | 40 秒内 10 个任务 | 4 个批次带检查点 |
集成测试: 所有 3 个技能自动串联,交付了可工作的 CLI ✅
示例工作流
用户: 构建一个 Markdown 笔记 CLI
→ 头脑风暴技能:
- - 提问:存储格式?需要搜索?需要标签?
- 提出:3 种方案(平面文件 vs SQLite vs JSON)
- 呈现设计,获得批准
- 保存:workspace/docs/plans/2026-02-25-notes-cli-design.md
- 调用编写计划
→ 编写计划技能:
- - 提问:TDD 还是直接实现?提交频率?
- 创建计划:24 个任务,每个 2-5 分钟
- 保存:workspace/docs/plans/2026-02-25-notes-cli-plan.md
- 调用执行计划
→ 执行计划技能:
- - 批次 1(任务 1-5):项目设置 → 检查点 ✓
- 批次 2(任务 6-10):创建笔记功能 → 检查点 ✓
- 批次 3(任务 11-15):列出笔记功能 → 检查点 ✓
- (继续直至完成)
结果: 可工作的 CLI 工具,完整文档、测试和内存追踪。
单个技能文件
- - brainstorming/SKILL.md — 279 行,完整流程 + 合理化计数器
- writing-plans/SKILL.md — 10KB,方法论问题 + 任务模板
- executing-plans/SKILL.md — 9KB,分批逻辑 + sessions_spawn 集成
每个技能可独立使用,也可作为完整工作流的一部分使用。
归属
基于 Jesse Vincent 的 obra/superpowers。
针对 OpenClaw 的个人助手架构进行了适配,包含内存集成、方法论问题(而非强制)、sessions_spawn 工作流和单一工作区模型。
许可证
MIT(遵循 obra/superpowers)
资源
- - GitHub: https://github.com/brothaakhee/superclaw(即将推出)
- 原始框架: https://github.com/obra/superpowers
- OpenClaw 文档: https://docs.openclaw.ai