Superpowers: Brainstorm, Plan, and True TDD for Claude Code

AI 一接到需求就埋頭寫 code，最後測試東漏西補？這套框架逼它先問清楚、先寫會失敗的測試、再寫剛好通過的 code。把資深工程師的紀律變成可貼上的工作流。

[ Log in to see the full prompt ]

## What This Is / What Problem It Solves Superpowers is an open-source "complete software development methodology" by Jesse Vincent (obra) that packages senior-engineer development discipline into composable "skills" fed to coding agents like Claude Code. It targets the four most common problems with AI-written code: ① diving into code before requirements are clear; ② vague specs where you only discover you're off-track halfway through; ③ insufficient testing, or tests bolted on after the fact; ④ a lack of systematic verification, with "should work" too often passing as done. Its core claim is: **don't jump into code the moment a request comes in** — instead, force the agent through a deliberate pipeline: spec → plan → test → verify. ## Why This Source Is Worth Using Its most valuable trait is how strictly it defines "true TDD": RED-GREEN-REFACTOR is mandatory (write a failing test first, then write just enough code to pass it, then refactor with the green light on), and it explicitly states that "code written before its test gets deleted" — a rule that shuts the door on the common practice of faking TDD (writing the code first and bolting tests on after). It also bakes in two practical principles, YAGNI and DRY, to keep the agent from over-engineering. The whole process is designed so the agent can "work autonomously for one to two hours straight without drifting from the plan." ## How to Use It (Steps / Key Points) The full_prompt above rewrites its spirit into a workflow instruction you can paste directly into Claude Code (or any long-context agent), in five stages: 1. **Brainstorm**: ask clarifying questions first, present the design in sections for your confirmation, and no writing code yet. 2. **Write a plan**: break the work into 2-5 minute tasks, each with an exact file path and a way to verify it. 3. **True TDD**: for each task, follow RED→GREEN→REFACTOR, pasting the test output at every step. 4. **Two-stage subagent review**: for multi-task work, dispatch a clean subagent per task — first to review spec compliance, then to review code quality. 5. **Wrap-up**: run the full test suite, summarize the changes, and offer merge/PR options. Before using it, just swap in your own content for {{FEATURE_REQUEST}}, {{TECH_STACK}}, and {{TEST_COMMAND}}. ## When to Use It Best for projects with "existing test infrastructure and a premium on correctness": backend services, libraries, modules with clear inputs and outputs, and refactoring existing code. Not suited to purely exploratory spikes, or throwaway prototypes where you haven't even decided to commit to the idea yet (use a lighter process for those). If you just want a quick single snippet generated, this framework will feel too heavy — what it buys you is less rework and verifiable correctness. ## Source + License 📎 Source: obra/superpowers (by Jesse Vincent, MIT License) — this entry is a Traditional Chinese adaptation; see the link above for the original content. The original repo provides a full skill system (test-driven-development, brainstorming, writing-plans, subagent-driven-development, using-git-worktrees, systematic-debugging, and more) and installation instructions for multiple harnesses; this entry only distills the spirit of its workflow into a ready-to-use prompt, without copying its code or skill files verbatim.