English | 中文
A universal skill pack that applies research-backed repetition techniques to dramatically improve AI agent accuracy and reliability. Works with Claude Code, Cursor, Windsurf, Continue, Aider, and any LLM-based agent system.
| Research | Finding | Impact |
|---|---|---|
| Google Research (2025) | Repeating prompts twice | +76% accuracy on non-reasoning tasks |
| Wang et al. (2022) | Self-consistency with majority voting | +30% on reasoning tasks |
| Brown et al. (2024) | Repeated sampling scaling laws | Small models beat large ones |
Transformer models process text left-to-right only. They can't "look ahead" to see what's coming. When you repeat a prompt:
- First pass: Model processes context without knowing the question
- Second pass: Model now knows the question and can properly attend to relevant context
It's like giving the model a second chance to connect the dots.
git clone https://github.com/systemime/repetition-enhanced.git
cp -r repetition-enhanced/skills/* ~/.claude/commands/
cp -r repetition-enhanced/commands/* ~/.claude/commands/git clone https://github.com/systemime/repetition-enhanced.git
# Copy SKILL.md content to your tool's custom instructions/rulesSimply apply the techniques manually when prompting, or import the skill definitions into your agent system.
# Double prompt for non-reasoning tasks
/boost double What is the 25th item in this list...
# Triple verification for math/logic
/boost triple Solve: If 3x + 7 = 22, what is x?
# Best-of-N for code generation
/boost verify Write a function to sort a linked list
# Full verification chain for critical tasks
/boost critical What are the security implications of...Double Prompt: Copy your prompt twice
[Your prompt]
[Your prompt again]
Self-Consistency: Ask for multiple solutions
Solve this problem using 3 different approaches.
Compare all answers and provide the majority result.
The skill automatically activates when:
- 🎯 Making important decisions (architecture, tech stack)
- 💻 Writing critical code (production, database operations)
- 🧮 Complex reasoning (math, logic, multi-step problems)
⚠️ High-risk operations (deletions, config changes, deployments)
Best for: Lookups, extractions, direct Q&A
Original: "Find the 25th name in this list..."
Enhanced: "Find the 25th name in this list... [REPEAT]"
Accuracy boost: Up to 76% (21% → 97% in research tests)
Best for: Math, logic, reasoning
Generate 3 independent solutions → Extract answers → Majority vote
Accuracy boost: 10-30% on reasoning benchmarks
Best for: Code generation, verifiable tasks
Generate N candidates → Verify each → Select best
Accuracy boost: 20-50% with sufficient samples
Best for: High-stakes decisions
Initial solution → Self-review → Alternative approach → Cross-validate → Synthesize
| Task Type | Technique | Repetitions | Cost |
|---|---|---|---|
| Simple lookup | Double Prompt | 2x | Low |
| Math/Logic | Self-Consistency | 3-5x | Medium |
| Code generation | Best-of-N | 10-100x | Higher |
| Critical decisions | Full Chain | Variable | Highest |
repetition-enhanced/
├── README.md # This file
├── plugin.json # Plugin configuration
├── LICENSE # MIT License
├── CHANGELOG.md # Version history
├── skills/
│ └── repetition-enhanced/
│ ├── SKILL.md # Main skill definition
│ ├── references/
│ │ └── research-papers.md # Academic references
│ └── examples/
│ └── patterns.md # Usage patterns
└── commands/
└── boost.md # /boost command
| Tool | Compatibility | Installation |
|---|---|---|
| Claude Code | ✅ Full | Copy to ~/.claude/commands/ |
| Cursor | ✅ Full | Copy SKILL.md to rules |
| Windsurf | ✅ Full | Import as custom skill |
| Continue | ✅ Full | Add to custom instructions |
| Aider | ✅ Partial | Manual application |
| Any LLM | ✅ Universal | Apply techniques manually |
- Prompt Repetition: Google Research (2025)
- Self-Consistency: Wang et al. (2022) - ICLR 2023
- Repeated Sampling: Brown et al. (2024) - ICLR 2025
Contributions welcome! Please read our contributing guidelines.
- Fork the repository
- Create your feature branch (
git checkout -b feature/amazing-feature) - Commit your changes (
git commit -m 'Add amazing feature') - Push to the branch (
git push origin feature/amazing-feature) - Open a Pull Request
MIT License - see LICENSE file for details.
Made with ❤️ by systemime
一个通用的技能包，应用学术研究支持的重复技术来大幅提升AI智能体的准确率和可靠性。适用于Claude Code、Cursor、Windsurf、Continue、Aider以及任何基于LLM的智能体系统。
| 研究 | 发现 | 影响 |
|---|---|---|
| Google研究 (2025) | 重复提示词两次 | 非推理任务准确率提升76% |
| Wang等人 (2022) | 自一致性+多数投票 | 推理任务提升30% |
| Brown等人 (2024) | 重复采样扩展定律 | 小模型击败大模型 |
Transformer模型只能从左到右处理文本，无法"向前看"。当你重复提示词时：
- 第一遍：模型在不知道问题的情况下处理上下文
- 第二遍：模型现在知道问题了，可以正确关注相关上下文
就像给模型第二次机会来建立联系。
git clone https://github.com/systemime/repetition-enhanced.git
cp -r repetition-enhanced/skills/* ~/.claude/commands/
cp -r repetition-enhanced/commands/* ~/.claude/commands/git clone https://github.com/systemime/repetition-enhanced.git
# 将SKILL.md内容复制到工具的自定义指令/规则中手动应用这些技术，或将技能定义导入您的智能体系统。
# 非推理任务：重复提示词
/boost double 在这个列表中找到第25个名字...
# 数学/逻辑：三次验证
/boost triple 解方程：如果3x + 7 = 22，x是多少？
# 代码生成：多样本验证
/boost verify 写一个排序链表的函数
# 关键任务：完整验证链
/boost critical 这个安全决策的影响是什么...遇到以下情况时自动激活：
- 🎯 重要决策（架构、技术栈）
- 💻 关键代码（生产环境、数据库操作）
- 🧮 复杂推理（数学、逻辑、多步骤问题）
⚠️ 高风险操作（删除、配置修改、部署）
| 工具 | 兼容性 | 安装方式 |
|---|---|---|
| Claude Code | ✅ 完全支持 | 复制到 ~/.claude/commands/ |
| Cursor | ✅ 完全支持 | 复制SKILL.md到规则 |
| Windsurf | ✅ 完全支持 | 导入为自定义技能 |
| Continue | ✅ 完全支持 | 添加到自定义指令 |
| Aider | ✅ 部分支持 | 手动应用 |
| 任何LLM | ✅ 通用 | 手动应用技术 |
- 提示词重复: Google研究 (2025)
- 自一致性: Wang等人 (2022) - ICLR 2023
- 重复采样: Brown等人 (2024) - ICLR 2025
MIT许可证 - 详见LICENSE文件。
Made with ❤️ by systemime