CLAUDE.md release based on Karpathy's observations improves Claude Code coding behavior
Decision Brief
The file comes from Andrej Karpathy's observations of LLM coding flaws, grouping the problems into four categories: the model makes wrong assumptions for you without checking, overcomplicates code and abstractions, and makes unrelated changes to things it does not understand. It proposes four corresponding principles: think before coding (state assumptions, present multiple interpretations, push back when necessary, stop when confused), simplicity first (write only the minimum code needed to solve the problem, no abstractions for one-time use, rewrite if 200 lines can be reduced to 50), surgical changes (touch only the code that must be touched, match existing style, do not delete pre-existing dead code), and goal-driven execution (turn instructions into verifiable success criteria and use a test-driven loop). There are two installation methods: the recommended one is to install it as a global plugin via the plugin marketplace within Claude Code, or per project download the CLAUDE.md file with curl; for existing projects, the content can be appended. The repository also includes a Cursor project rules file that takes effect when the project is opened. For developers using Claude Code, this set of constraints is suitable for merging into an existing CLAUDE.md and project rules, especially for projects with strict code review and sensitive change scope; however, the file explicitly states that it favors caution over speed, so trivial tasks like fixing typos or one-line changes do not require full execution.
Sources
- Skill Radar(GitHub 趋势)
Trending hands-on MCP servers, agent skills, and AI-coding tools discovered from GitHub search momentum.
- GitHub:multica-ai/andrej-karpathy-skills
留言
登入后即可留言,和其他 builder 交换实测心得。
还没有留言,抢头香。