Tips for Using AI Coding Tools
A few practical tips
A Few Useful Prompts
When you think it may not understand your request, such as when you feel your description is unclearPlease restate my request
After asking it to debug several times without resultsPlease add detailed logs at every relevant location in the code
When you are dissatisfied after several implementation attempts, or when starting a debugging session and want it to clarify its logicPlease explain your implementation logic immediately
When you already have a successful similar feature and want it to implement a similar onePlease refer to the implementation logic used elsewhere
Development Advice
- If the project is important, be proactive about review and refactoring guidance based on your own preferences. Otherwise, the code will become increasingly bloated, the blind spots you are unfamiliar with will grow, and more redundant code will accumulate, such as duplicate implementations or unused logic. This can create gaps in understanding when you and the tool develop new features together.
- Actively use Claude Code plan mode for new feature planning, brainstorming solutions, and roadmap planning.
- The hardest cases to debug are those where logs cannot be printed, especially frontend interfaces. They are also difficult to describe when the interface is complex.
Use Case Analysis
There are two ways to use it: for work projects and for personal projects.
Personal Projects: Building a Product from 0 to 1
If you want to build a personal product from 0 to 1:
- Gemini 2.5 Pro -> requirements document and technical architecture document;
- Then Claude Code (Claude 4.5 Sonnet) -> pass the requirements document and technical architecture document to Claude Code and have it convert them into a development plan;
- Put these three files in the project folder;
- Then ask Claude Code to execute the development plan and update the files.
Work Projects: AI Enablement
If you want to use AI to improve your work:
- In the past, I would have recommended Cursor, but I no longer recommend it because Cursor has been very unstable recently. This instability is inevitable. The core reason is that Claude Code comes from a primary LLM provider. With their $20 membership, you can use roughly $100 worth of value, but Cursor, as a second-level AI company, lets you use at most a little over $40, even with roughly double overage. I still recommend Claude Code.
Development Practice Suggestions
- Before some larger commits, you can let it finish writing the code, then ask it to self-check for security issues, quality issues, and redundancy, and then review it manually.
- Do not try to do too much at once. Let it solve bugs one by one and implement features one by one. This keeps the context focused and increases the success rate.
- Running Claude Code on two projects at the same time works well. Running three at once makes it harder to focus and is quite tiring.
- If AI cannot solve something, you can also try searching on Google. You might be pleasantly surprised.
Tool Recommendations
- Claude skills are a good thing and very useful. The more useful ones include feature-dev and frontend-design.
- Among MCPs, the official Figma MCP from Figma is quite practical. It is suitable for companies that use Figma for design, and it makes frontend implementation from designs very convenient.
Some Other Thoughts
- Claude Code implements features very quickly, but in reality it shifts the pressure to testing. Most of the time, I am not a full-stack engineer but a tester, haha, or even an operator.
- AI is powerful, but for now it still cannot replace us in building real interpersonal relationships, emotional connections, trust, and similar things.
Comments
- Every time I finish describing a requirement, I add this sentence afterward: "At this stage, do not generate any code. Is the problem description clear? Your responsibility is to perform planning and raise any unclear or questionable parts of the task."
- Can frontend work be handed over to Gemini 3 Pro?
- That would be better, and I recommend it. I use Claude frontend-design for frontend design. The main reason is that I am too lazy to switch models and tools.
- I have been using Gemini 3 over the past few days. Compared with Claude: Advantages: fast speed; low price; good generated code quality; Disadvantages: gemini-cli is not as easy to use as Claude Code; instruction following is relatively poor. For example, when I ask it to use Chinese or describe the modification plan first, it often uses English or writes code directly. I mainly write Python, Java, and Vue.
Reference Sources
Some Experiences After Using Claude Code So FarQ&A