Claude Code has completely changed the way we develop our app, Chorus, so much so that we now consider it one of our employees. At $200/mo. it’s earning far below minimum wage, and we only put it on PIP once.

We now put an asterisk on the whiteboard next to any task that Claude will mostly handle. And the asterisks are multiplying.

Here’s what we’ve learned about how to manage Claude.

Get the easy stuff right

Run /model to see which model you're using. Set it to Opus. It’s more expensive but worth it. I accidentally had it on the default (Sonnet) for weeks. I thought Claude was quiet quitting but I was just using his dumber, cheaper cousin.

Use the Cursor/VSCode extension. You now have a shortcut (CMD+Esc) and it automatically reads the file you’re in and any text you’re highlighting. I’m still using Cursor, but the Claude Code extension has almost completely replaced the AI chat sidebar.

Start with plan mode

(Shift+Tab Shift+Tab to activate)

This is a recent addition and I’ll use it for any work that doesn’t have a clear and small scope. I have no hard data but planning seems to improve results, and it also helps me think through what I want.

Like any AI, Claude loves to overcomplicate things. It’s a token fountain. Every line of code it writes should give you pain.

I will usually remind it to “keep it as simple as possible.” We don’t need to solve every edge case, we’re a startup, you fool!

Feed the beast

Paste in links. Paste in images. To paste images you have to use ctrl+v (that’s right, ctrl, not cmd!). I had no idea this existed until we wrote this post and Jackson told me.

Claude can use git and the GitHub command line tool, gh. Ask it to make PRs, resolve merge conflicts, debug using git bisect, etc.

Don’t expect it to know what files to look at. Guide it with some @ s. Claude is good at searching but not amazing, it needs your help.

Imagine you’re sending a human engineer your request. Except you don’t have to be as polite (but maybe you should, who knows).

Here’s a tip from Omid, a human: “I’ve found the more I talk to Claude like an engineer I’m pair programming with, the better it performs. This has helped me get over the hump of figuring out what context to give it. You can act like you’re writing a Jira ticket for a junior engineer.”

Use the magic word(s)

I also will often add “ultrathink” if I want a bit of extra oomph. This is like a secret magical spell to tell Claude to use more tokens. It’s bizarre to me that we live in a world where I am telling you magic secret passphrases to whisper to your AI to make it code better for you, but here we are.

Watch it cook, or not

Once it’s made a plan, tell it to start implementing. I’m a trusting type so I always Shift+Tab to auto-accept edits.

Don’t be afraid to esc and do some more explaining if you think it’s going down a bad path.

Something I still haven’t figured out: what do I do while Claude is working? It’s impossible to get into a flow state with these vibe coding tool like I do when I code with my raw brain. Here’s one solution, let me know if you have any better ideas.

🚨 The Most Dangerous Trap 🚨

If you are having trouble—let’s say you’ve asked Claude the same things 3–4 times with different strategies and you’re getting nowhere—pull the eject button and give up. You are entering what I call the “rabbit hole problem.”

Vibecoding is powerful but the absolute worst version of it is lazily throwing stuff at the wall, it not working, and then spamming the chat. It's really hard to fight this, and I usually catch myself when it’s too late, when I’m tired and I don’t want to put in the effort of actually thinking through the problem.

Something that helps is to think of all the work as helping YOU think about the problem. Don't view code the AI wrote as having the same weight as code you wrote. You can throw it all away and it doesn’t matter.

One useful strategy here is to start over with a fresh context (see “Ask Claude to review” below).

Review the code

The hardest part is knowing whether Claude’s work was any good.

With Chorus, Claude can’t interact with our app, so I ask it to write a testing plan I can manually check off. Here’s a testing plan Claude wrote in a PR description (as you can see, I skipped the hard parts):

You can also spin up a second Claude instance and ask it to review the work. Each Claude tends to be very attached to its own work, and will find it much easier to criticize other Claudes. (Hey, that reminds me of someone!) Often I’ll ask some other AIs in Chorus if I feel like I’m going down a bad path.

How do you spin up a new Claude instance? The easy way is /clear, but that will also irrecoverably wipe out your chat context, which you might want to return to. I usually just open a new pane in my terminal. Yes, this eventually becomes a huge pain.

On harder tasks, I’ve found myself repeating, at least 2-3 times: Can you double check this is as simple as it can be?

As a last step, and I know this sounds insane, but hear me out—read the code yourself. Be suspicious. Claude loves to cut corners. It’s constantly using ref to get out of changing the React state model, or using typecasts so it can avoid figuring out the right types.

Keep teaching

Whenever Claude messes, put # in front of the correction. The instruction gets saved to Claude’s memory (stored in CLAUDE.md), and Claude will follow it in future sessions.

If possible you should check CLAUDE.md into source control.

When not to use Claude

I don’t use it for UI work. There’s a distinctive LLM design language that I can’t quite put my finger on but just looks shitty. I see it in apps created with Lovable, Bolt, etc. Quintessential example:

Human designers (like Julian) are miles better.

I also wouldn’t use Claude to write code I’m going to build on a lot later on, where each tiny design decision I make now will become a massive binding constraint down the road. It’s still helpful for thinking through how to implement these things, though.

What does this all mean?

Things are changing really fast, and we’re still figuring out how to work with our newfound sand intelligences.

I’ve been surprised how much I’ve started to trust Claude. The game changer here is that it can do full scoped units of work. Even for things it can't finish, it gets the ball rolling, which is often the hardest part with coding.

Congrats to you, Claude, employee of the month!