How I Work with Claude Code (and Why You Might Want To)

Lessons from a couple of months pairing with a GenAI CLI assistant

1. Don’t start by asking it to code. Start by thinking together.

Claude Code isn’t just an AI that types fast.
It’s an AI that thinks with you—if you let it.

When I began building a lead enrichment dashboard, I didn’t ask it to create buttons or write reducers.

I asked:

“What are the most important widgets for this problem?”

It had opinions. Some sharp, some overkill. But it made me think better. So I started adding a line to most prompts:

“Don’t write code—let’s just chat.”

---

2. Your terminal becomes a collaborative space.

This thing lives in the CLI, not your browser.

And that matters. It means you can:

• Prompt on the left

• Approve code in your IDE on the right

• Work without toggling contexts or drowning in GUIs

• This also means you can write prompts like "install rust" and it will

Planning mode keeps it conversational until you’re ready to build.
And every edit gets a diff review. You stay in control.

---

3. Set expectations clearly. Literally.

Claude reads a claude.md file in your project.

Mine includes:

• Don’t change more than one thing at a time

• Work test-first

• Ask before writing

• Prefer clarity over cleverness

Every few days, I ask:

“What have you learned about how I like to work?”

Claude answers. I copy the useful bits into claude.md. That file is now the onboarding doc I wish human teammates read.

---

4. It runs tests. Then learns from them.

This surprised me.

I’d say:

“Write a failing test. Then make it pass.”

Claude would:

• Write the test

• Run it

• Read the error

• Fix the code

• Repeat

It’s iterative dev, powered by a large language model, at CLI speed.
And lately? It gets it right more often than not.

---

5. The code is getting better. So is the assistant.

When I started, Claude Code had rough edges. Missing imports. Subtle bugs. Broken setups.

But it improved. Fast.
Now I rarely see compilation or runtime errors.

It’s not just me adapting to the tool—it's the tool adapting to the work.

---

6. Starting something new doesn’t feel heavy anymore.

Before Claude Code:
New framework = a long weekend of setup and tab-hopping.

Now:

“Create a Rust project” → done.
“Spin up a React UI” → done.
“Here’s an example—learn this style” → done.

It doesn’t just copy patterns. It learns structure, tone, and architectural choices.
Taste matters. Claude gets that.

---

7. It reads the docs. So I don’t have to.

I give it a link to documentation. Then I say:

“Use this library to write X.”

It does. Correctly.
No summarizing. Just understanding and applying.

It’s like onboarding a new dev who actually reads the docs.

---

8. What works (for me):

• Work in tight loops

• Ask before code

• Use claude.md to define your rules

• Gate changes via IDE diff

• Stay test-first

• Feed it your best examples

• Let it understand, not summarize

---

9. Where this is going

Claude Code is already better at writing correct code than it was two months ago.
What’s next?

Writing well-crafted code.
Not in years. In months.

---

10. Why I stuck with it

Because it made me think more clearly, explore faster, and debug with less pain.

And why Claude Code, because I prefer it. May be you can do this with many other tools, idea is to find your Bodhi Tree and sit under it!

Thanks for reading till here, if you were looking for some of the rules file I have been using, you can find them here.

---