10 things I learned in a month with Claude

After a month with Claude I have a notebook full of notes. The full story of how I got into AI — from the free tier through counting ducks to carfast.cz — I described in the article From free tier to dependency. Here are the concrete tips that actually work for me.

1. Sonnet for everything, Opus for thinking

You don't need to use the most expensive model. Sonnet 4 handles 90% of tasks — it's fast, cheap, and smart enough. I reach for Opus only when I need to:

• Analyze a large codebase (150+ files)
• Do deep reasoning over architecture
• Find a subtle bug in complex logic

Rule: start with Sonnet. Switch to Opus only when Sonnet isn't enough.

2. Context is king

The more context you give Claude, the better the answer you get. Don't write "fix the bug." Write:

V souboru /src/api/checkout.ts na řádku 47 padá TypeError:
Cannot read property 'price' of undefined. Funkce se volá
z CartComponent. Tady je relevantní kód: [kód]

Claude isn't a mind reader. It's a colleague — and a colleague also needs context.

3. CLAUDE.md changes the game

Create a CLAUDE.md file in the root of your project — project description, conventions, tech stack. Claude Code reads it at the start of every session. Instead of repeatedly explaining "we use Tailwind, not CSS modules," you write it once.

My CLAUDE.md for airepublic.cz has: tech stack, directory structure, naming conventions, content tone. 15 lines that save hundreds of words in prompts.

4. Iterate, don't restart

Bad workflow:

1. Write prompt → get reply → not quite right → rewrite the whole prompt

Good workflow:

1. Write prompt → get reply → "change X" → "add Y" → "remove Z"

Claude remembers the conversation context. Use it. Iteration beats restart.

5. Turn extended thinking on selectively

Extended thinking is great for complex problems, but for a simple "rewrite this function" it's overkill. It slows things down and burns more tokens.

I enable it for:

• Architectural decisions
• Debugging mysterious errors
• Comparing several approaches
• Security risk analysis

6. Claude Code > copy-paste from chat

The biggest productivity jump wasn't better prompting — it was moving from the web chat to Claude Code in the terminal. The difference:

• Chat: write → get code → copy → paste → edit
• Claude Code: write → the code is in the file → test → commit

It's not just convenience. It's a fundamentally different workflow — Claude works with your project, not beside it.

7. MCP servers unlock superpowers

MCP (Model Context Protocol) connects Claude to external tools. My setup:

• GitHub — reads issues, opens PRs
• Filesystem — works with files outside the repo
• Vercel — deploy status, logs

Set it up once, it works forever. Claude goes from a clever chatbot to a tool wired into your ecosystem.

8. Limits are real — plan around them

Claude Pro has limits. Under heavy use you'll hit them. My strategy:

• Break tasks up — instead of one giant prompt, split into smaller steps
• Multiple sessions — in Warp I have 2-3 Claude Code tabs for different tasks
• Batch work — group similar tasks together
• Take a break — when you hit the limit, grab a coffee. Your brain needs it too

9. AI doesn't write perfect code on the first try

Claude writes working code. But not always optimal code. Always:

• Read what it gave you back (even if it "works")
• Test edge cases
• Check security (AI sometimes forgets to sanitize input)
• Refactor if you see a better path

Claude is a junior developer with encyclopedic knowledge. Your senior judgment is still required.

10. Save the prompts that work

When you find a prompt that consistently produces good results, save it. I have a prompts.md file with templates for:

• Code review ("go through this PR, focus on...")
• Refactoring ("rewrite as a Server Component, keep...")
• Debugging ("analyze this error, context is...")
• Content ("write an article in this tone, structure is...")

Prompt engineering isn't a science — it's a collection of tips that work for your specific use case.

Bonus: don't expect this

• AI won't replace programmers. It'll replace programmers who don't use AI.
• You won't be faster right away. The first week is learning. From week two you're flying.
• You'll think differently. Instead of "how do I code this," you start asking "what do I want it to do." And that's a better question.