本文由 AI 分析生成
建立時間： 2025-10-09
Summary
EN: Jesse describes his “Superpowers” Claude Code plugin, which implements a SKILL.md-based system for AI-assisted coding. Key workflows: brainstorm→plan→implement pipeline, RED/GREEN TDD (AI writes failing test first, then implementation), git worktrees for parallel experiments, subagent dispatching for parallel work, Cialdini-based persuasion principles embedded in SKILL.md, and a SQLite vector index for persistent memory across sessions.
ZH: Jesse 描述他的「Superpowers」Claude Code 插件，實作以 SKILL.md 為基礎的 AI 輔助編碼系統。主要工作流：腦力激盪→規劃→實作管線、RED/GREEN TDD（AI 先寫失敗測試再實作）、git worktree 平行實驗、子代理平行工作分派、SKILL.md 中嵌入 Cialdini 說服原則，以及 SQLite 向量索引實現跨會話持久記憶。
Key Points
- SKILL.md system: custom skill files that give Claude specific context about project conventions, patterns, and constraints
- Brainstorm → Plan → Implement: explicit three-phase pipeline; AI generates options before committing to an approach
- RED/GREEN TDD: AI writes a failing test (RED), then writes implementation to make it pass (GREEN) — forces explicit specification before implementation
- Git worktrees: multiple working directories on the same repo; enables parallel agent experiments without branch switching
- Subagent dispatching: break large tasks into parallel sub-tasks assigned to separate agent instances
- Cialdini principles: authority, social proof, reciprocity embedded in SKILL.md to influence Claude’s reasoning toward better outputs
- SQLite vector index: persistent memory that survives across Claude sessions; enables context without re-explaining
Insights
- The Cialdini persuasion principles in SKILL.md are unusually creative: treating the AI like a persuadable agent whose behavior can be shaped by social/authority signals rather than just instructions
- Git worktrees + parallel agents is an underused pattern — most developers run agents sequentially when many tasks are independent
- The RED/GREEN TDD pattern enforces specification discipline on the AI: it can’t cut corners by writing implementation without a clear test
Connections
- Directly implements the Skills architecture Armin Ronacher advocates for in the skills vs MCP article
- SkillsBench empirically validates the SKILL.md approach: curated, focused skills +16.2pp
- The git worktrees pattern appears in the Claude Code best practices article in this vault
Raw Excerpt
“The RED/GREEN pattern forces the agent to specify what success looks like before touching implementation. First, write a failing test that precisely describes the desired behavior. Then make it pass. It sounds simple but it eliminates an entire class of AI-generated code that looks right but doesn’t actually work.”