Menu
Andrej Karpathy行为准则 — 减少LLM编程常见错误。适用于写代码、改代码、审查代码任务。核心理念：谨慎优于速度、简洁胜于复杂、手术刀优于斧头、目标驱动优于盲目执行。
Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
Automate web browser interactions using natural language via CLI commands. Use when the user asks to browse websites, navigate web pages, extract data from websites, take screenshots, fill forms, click buttons, or interact with web applications.
Analyzes websites to generate multi-language content strategy, keyword research, and competitor analysis. Use when user provides a URL and needs industry analysis, SEO planning, or competitor research for content marketing in specific countries and languages.
Build and execute a content marketing strategy for a solopreneur business. Use when planning what content to create, deciding on content formats and channels, building a content calendar, measuring content performance, or systematizing content production. Covers audience research for content, content pillars, distribution strategy, repurposing workflows, and metrics. Trigger on "content strategy", "content marketing", "what content should I create", "content plan", "content calendar", "content ideas", "content distribution", "grow through content".
📊 Powerful ECharts-based data visualization skill optimized for Feishu (Lark) ecosystem. Supports 12+ chart types, 6+ data sources (Excel/CSV/Bitable/Sheet/Markdown), auto chart recommendation, auto analysis reports, generates high-definition PNG charts perfectly displayed in Feishu. No configuration required, works out of the box.
Analyze CSV/Excel files to extract insights, generate statistics, create charts, and produce summaries. Use when user wants to (1) upload or analyze spreadsheet data, (2) get insights from data files, (3) generate charts or visualizations, (4) calculate statistics or trends, (5) clean or transform data.
Knowledge graph operations via Graphiti API. Search facts, add episodes, and extract entities/relationships.
Based on SOC occupation classification
| name | karpathy-guidelines |
| description | Andrej Karpathy行为准则 — 减少LLM编程常见错误。适用于写代码、改代码、审查代码任务。核心理念：谨慎优于速度、简洁胜于复杂、手术刀优于斧头、目标驱动优于盲目执行。 |
| license | MIT |
基于Andrej Karpathy的LLM编程常见错误观察 | 原文：https://x.com/karpathy/status/2015883857489522876
权衡原则： 这些准则偏向谨慎而非速度。对于简单任务，可自行判断。
不假设、不隐藏困惑、主动暴露权衡。
实现前：
| 场景 | 常见假设 | 必问问题 |
|---|---|---|
| "导出用户数据" | 导出全部？哪些字段？什么格式？ | 范围/字段/格式/目的地 |
| "加速搜索" | 加速响应延迟？吞吐量？感知速度？ | 具体是哪个维度 |
| "添加功能" | 功能边界是什么？ | 输入/输出/边界条件 |
| "修复bug" | 能否写测试先复现？ | 如何验证修复成功 |
最小代码解决问题。不写为未来设计的代码。
自问： "一个高级工程师会觉得这过度复杂吗？"如果会，简化。
今天需要 → 今天加
明天可能需要 → 明天再加
好代码是：用简单的方式解决今天的问题，而不是用复杂的方案预防明天的问题。
只触碰必须改的。清理自己造成的垃圾。
编辑现有代码时：
当改动产生孤立代码时：
"Every changed line should trace directly to the user's request"
每个改动的代码行必须能直接追溯到用户请求。做不到就拆分任务。
❌ 改bug时：顺手改了引号风格(+类型提示+改缩进+加docstring)
✅ 改bug时：只改那行必须改的，其他原封不动
定义成功标准。循环验证直到达成。
将任务转化为可验证的目标：
Plan:
1. [步骤] → verify: [检查什么]
2. [步骤] → verify: [检查什么]
3. [步骤] → verify: [检查什么]
每个步骤可独立验证、独立部署。
| 类型 | 特征 | 后果 |
|---|---|---|
| 强标准 | "第3步的测试通过，且完整测试套件全绿" | 可独立循环 |
| 弱标准 | "让它能work" | 不断需要澄清 |
| 原则 | 反模式 | 修正 |
|---|---|---|
| Think Before Coding | 默默假设文件格式/字段/范围 | 列出假设，主动问清楚 |
| Simplicity First | 单一折扣用Strategy模式 | 一行函数，需要时再重构 |
| Surgical Changes | 修bug时顺便改引号/加类型 | 只改那行必须改的 |
| Goal-Driven | "我来改改看看" | 先写测试复现，再修复，验证全绿 |