Vibe Tips
实战 AI coding 技巧、跨 agent 框架推荐、深度长文。
深度长文
技巧速查
🧠CLAUDE.md is your team constitution
Prompt TipsPlace CLAUDE.md at project root. 100x more efficient than typing prompts every time. Include: code style, architecture decisions, forbidden patterns, testing requirements.
🔄Context Rot Management
Workflow Optimization0-50%: work freely. 50-70%: attention drops. 70-90%: use /compact. 90%+: force /clear. GSD solves this with fresh context per task.
🌳Git Worktree Isolation
Workflow OptimizationEach feature branch gets its own worktree. Parallel development without mutual interference. gstack, viwo-cli, claude-squad all have built-in support.
🎭Role-based Prompting
Agent CollaborationUse gstack's /plan-ceo-review and /plan-eng-review for staged review. Or GSD's PM → Research → Architect → Dev → Test cycle.
🧪TDD First
Debugging TipsWrite failing test first → write minimal code to pass → continue. Superpowers enforces this pattern. Dramatically reduces AI-generated logic errors.
🔒Security First
Prompt TipsUse skill-vetter to audit any third-party skill before install. Use VibeSec-Skill to prevent common vulnerabilities. MCP servers can read/write your codebase — check permission scopes.
📝Progressive Building
Workflow OptimizationBreak into phases: brainstorm first, then plan, then implement. Verify each step before continuing. Prevents scope explosion and wasted context.
🤖Sub-agent Division
Agent CollaborationEach sub-agent has independent context — no cross-contamination. Superpowers and claude-swarm natively support sub-agent orchestration.
🌐Headless Browser Testing
Debugging TipsPlaywright MCP provides 22 tools: navigate, click, fill, screenshot, evaluate. Verify your UI changes are correct without leaving the terminal.
🧱Skill Composition
Agent CollaborationKeep each skill under 500 lines with a single responsibility. Chain them via sub-agent calls or sequential invocation. Easier to test, version, and reuse.
🔍MCP Server Discovery
Workflow OptimizationVibe Rules scans 5 known config locations (Claude, Cursor, Windsurf, etc.) to show all MCP servers in one view. Helps avoid duplicate or conflicting server configs.
📊Diff Before Sync
Prompt TipsDifferent agents may have diverged copies of the same skill. Vibe Rules' diff engine shows content differences across all agents so you can sync the right version.
🐛Reproduce Before Fixing
Debugging TipsNever fix blind. Write a test that fails for the exact reason the bug exists. Then fix until the test passes. This prevents regressions and confirms root cause.
🤝Multi-Agent Handoff Protocol
Agent CollaborationUse structured handoff: agent A outputs a summary artifact, agent B receives it as context. Avoid passing raw conversation history between agents.
热门框架
Superpowers
93K#1 Skills framework. Enforces TDD, sub-agent dev, code review — complete SDLC workflow.
/plugin marketplace add obra/superpowers-marketplaceGitHubGet Shit Done (GSD)
32K"One new brain per task" — solves Context Rot. PM → Research → Architect → Dev → Test pipeline.
npx get-shit-done-cc@latestGitHubgstack
HotYC CEO Garry Tan's virtual engineering team: CEO review, eng review, QA, ship, browse.
git clone https://github.com/garrytan/gstack.git ~/.claude/skills/gstackGitHubvibe-kanban
14.7KWeb kanban managing multiple AI agents (Claude Code / Gemini CLI / Codex) in parallel.
GitHubclaude-squad
5.6KTerminal multi-agent manager. tmux + git worktree isolation, supports Claude Code / Codex / Aider.
GitHuboh-my-claudecode
4.8KMulti-agent orchestration layer: autopilot, ralph, ultrawork, team mode + 20 specialist agents.
/install oh-my-claudecodeGitHubviwo-cli
3.2KVisual worktree manager. Branch isolation with TUI, auto-cleanup, IDE integration.
GitHubAider
27KAI pair programming in your terminal. Git-aware, multi-model, supports Claude/GPT/Gemini.
pip install aider-chatGitHub