Impressions on Using AI Programming Tools

Aug 2, 2024

AI Programming Tools Tested

ChatGPT: Best suited for small-scale code iterations.
Claude / Claude Code: Excellent for small project modifications. When paired with codemcp, it allows direct local code editing — essentially a free alternative to Cursor or Claude Code.
Cursor: The overall results were unsatisfactory; it didn’t meet expectations well.
Replit: Great for small prototype websites. Deployment is one-click, but it’s not suitable for building websites with complex features.

After comprehensive testing, Claude Code demonstrated the strongest programming ability, while Replit offered the best UI.

Almost zero debugging ability: It doesn’t actually execute code, observe errors, or perform step-by-step debugging. It works through language prediction (predicting the next plausible token) to simulate code logic. This means it only infers solutions based on your provided context and error messages — it does not actively run code for trial and error, nor does it “inspect variables” through real debugging. Essentially, an LLM predicts rather than understands.
AI excels at “draft generation” but struggles with “final refinement.”
Why AI performs poorly in visual design: Visual aesthetics blend intuition and logic. AI currently handles only the logical templates, while beautiful UI requires visual grammar — such as whitespace, contrast, and balance — not linguistic grammar.
AI’s breakthrough in aesthetics won’t come from a single LLM but from the synergy of perception, comparison, expression, and evolution (reinforcement learning, RL). The future truly “aesthetic” AI won’t just “talk” — it will see, compare, choose, and refine like a multimodal intelligent agent.

Always paste the full error message when problems occur.
Clearly state your goal to the AI.
Remember, AI doesn’t understand you — it merely guesses what you might want. The more it guesses, the more likely it makes mistakes.
It doesn’t truly understand what the code is doing; it only predicts the most reasonable next snippet from context.
AI often writes code that’s too complex, verbose, or overly safe — you’ll need to manually trim and simplify it to match your own style.

Improve your questioning ability — through language learning: Read more philosophy and linguistics books to enhance how you express and structure prompts.
Enhance your aesthetic sense — through visual learning: The human brain naturally extracts patterns from large amounts of visual input, even without formal training. Through perceptual learning, the more excellent designs you observe, the sharper your sense becomes. Frequently browse platforms like Dribbble, Behance, or well-designed brand websites — your internal reference frame will rise naturally. Whatever you feed your brain, it will reproduce. The more you see, the more judgment you develop; the more you refine what you see, the more aesthetic sense you build. Even if you can’t explain why, you’ll know instinctively “what feels right.” That’s the power of perceptual learning.