Author: Wang Jianshuo
A simple record of my experiences with Claude Code so far, purely personal exploration, not necessarily applicable to everyone.
1. Stick to one tool and use it heavily. I use Claude Code. I don't necessarily think it's better than Codex, but the ROI of comparing tools may not be high, even though talking about the differences in detail gives a false sense of achievement.
2. Remember the most important keyboard shortcuts. Control+G to open the editor, helpful for writing longer content; Control+A, Control+E, Control+U—these are very practical shortcuts for quickly moving the cursor in the command line. Although not new to the AI era, they are as important as Control+C and Control+V when using.
3. Use voice input. HoldSpeak is very helpful.
4. Start a project by writing a PROJECT.md, using a structured method to write down everything you think of at once.
5. Claude agents are the default mode.
6. Claude Code, github.com, and cloudflare.com are a perfect match. Hand over the build process, release process, and all domain-related operations to the infrastructure.
7. Separate what is written by humans and what is written by machines. Manually maintain the core CLAUDE.md. Don't read the .md files or code written by Claude Code. Let machines do machine things, humans do human things. Understand what AI writes by asking the AI; don't look at the source code.
8. Drag and drop files into the Claude Code window—audio, video, documents, screenshots. If you can't explain something clearly, use Command+Shift+5 to take a screenshot and drag it over—fastest method.
9. Reconstruct the memory system. Centered on ~/.claude/CLAUDE.md, categorize and reference multiple memory files. Require not using the project's memory, and place all memory files in git, synchronized to GitHub (private). This way, your memory is permanent and cumulative, not scattered across each project.
10. Write Skills, and at the end of each work session, ask Claude to "precipitate what was learned into Skills"—this can be done automatically.
11. Whenever possible, for complex tasks, use ultracode to trigger dynamic workflows. Although expensive and slow, the results are guaranteed.
12. Accumulate Skills along the way, and continuously refactor Skills. Skills need to be placed in git.
13. Use git documentation as the output of the previous task and the input of the next task. Let agents have clear handover documents, without relying on context for connection.
14. Treat Claude Code as a horse (or a person), not as a car. A car turns under your direct command, a horse has its own ideas; we only need to set goals and boundaries. Its autonomous pathfinding characteristic is a feature, not a bug.
Does anyone have more to add?
