While Claude's been around for a little less than ChatGPT, you wouldn't really know it from its fanbase alone. It went from a somewhat niche tool only developers were following to a tool every other person is leaning on. Suddenly, everyone is using Claude and Claude Code to automate their workflow, ship the gazillion product ideas they've always had, and generally do more than they ever thought they could on their own. Claude Code is one of Anthropic's flagship products, and while that too once only found traction within a developer's workflow, it's now creeping into the workflows of designers, founders, marketers, and plenty of people who'd never call themselves engineers.

That said, a lot of Claude Code's users use it the fairly basic way. They open the terminal, run the claude command, and prompt it like any chatbot. What they miss is that Claude Code has a lot of layers that run much deeper. And while sure, power users love sharing their setups (we're guilty of it too), who better to listen to than the team that builds Claude Code every single day? Fortunately, the team behind Claude Code shared a few tips and tricks at Code with Claude, Anthropic's developer conference.

The CLI is just one of three ways to run Claude Code

Pick the surface that actually fits what you're building

Whenever I mention Claude Code to someone who hasn't used it before, their mind immediately defaults to a terminal, and the classic movie-hacker image of cryptic green text scrolling down a black screen. It looks intimidating, and for a lot of people, that alone is enough to keep them away. What a lot of people don't realize is that the terminal is just one of three ways to run Claude Code.

During the developer conference, Anthropic shared that it's expanded the number of ways you can use Claude Code. It boils down to three ways: the CLI, IDE, and Desktop. Each one is built for a different way of working, and the trick is matching the surface to what you're actually doing rather than forcing everything through the terminal. The CLI is what arrived first, and Anthropic revealed in the conference that it's still the interface that power users who want a minimal text interface and the most control (with the most customization features) prefer.

Then came the IDE, which gives users the same powerful agents and capabilities but with the ability to follow along with the code changes it's making, right there in a UI. And finally, Desktop, which is the newest addition and the most approachable of the bunch. It's a full-screen app with previews, images, and rich outputs. It also lets you manage all your agents, and gives you control over both your local and remote sessions.

So if you've been living in the terminal because you assumed that was Claude Code, that's the first habit worth breaking. The smarter move is to stop defaulting to one surface and start matching it to the task: reach for the CLI when you want raw control, the IDE when you want to watch the changes land, and Desktop when you want to manage everything at a glance.

Set up Routines and let Claude work while you sleep

Touch grass while Claude touches the codebase

Boris Cherny, the creator of Claude Code, has been very vocal about his setup. I've talked about his setup tips and tricks, and he's never been shy about admitting that Claude Code writes the overwhelming majority of his code these days. Now, the fact that Claude Code writes most of his code for him hasn't changed, but how that code gets written has. He's moved from prompting Claude one task at a time to setting up Routines and letting them run, often without him watching at all.

He describes them as "higher-order prompts," and explains that developers can set up async automations with Routines and wake up to ready-to-merge PRs. They can be run on a schedule or kicked off via webhooks or arbitrary API calls. They can also be run remotely or locally on your computer. This is also where that Desktop surface from earlier earns its keep — it's built for managing exactly this kind of async, multi-session work, so you can see which Routines are running and which need your input at a glance.

Let Claude check its own work

"Claude prompting Claude" is the new default

This is something Cherny has touched on before (which means I've also written about it): the idea of giving Claude a way to check its own work. Back then, the advice was to hand Claude a verification path, like the Claude in Chrome extension so it could see your frontend, or Computer Use so it could test an app, and let it confirm things actually worked.

At Code with Claude, that same idea showed up as the new default. Instead of you prompting the check, the agent now writes the code, triggers its own testing, and verifies the result before it ever surfaces a finished PR. The clearest example Cherny gave during the conference is CI auto-fix. Claude watches the PR the prior session just opened, and it then babysits it all the way to production. As Anthropic put it on stage, the person who owns the PR is never going to see a red X. You're no longer the one nudging Claude to go fix the failing check. Instead, it's watching and fixing on its own. It's the same principle as that old tip, just automated end to end.

The basic move is to write code and hope it works. The one the team actually leans on is to set things up so Claude can see the results and correct itself, then let it. It honestly feels a little dystopian, but Anthropic is framing this as having Claude prompt Claude, rather than you sitting there prompting it every step of the way.

Tell Claude what the final outcome actually looks like

Stop prompting and praying

One of the fastest ways to get a mediocre result out of Claude Code is to be vague about what you actually want. If you don't give Claude Code a clear outcome from the get-go, you'll end up constantly saying stuff like make it better, clean this up, add tests, and so on. Unfortunately, these give Claude nothing concrete to aim for, so it guesses, and you end up going back and forth.

At Code with Claude, the team made the case for the opposite approach: define what success looks like up front, in clear, testable terms, and let Claude work toward that target. It announced a feature called Outcomes at the conference, where you describe what success looks like and Claude iterates until it gets there. This is part of Claude's managed agents, and it's currently in public beta. That said, a short, simple markdown file where you spell out your criteria does the same job!

The best tips come from the team itself

There's no shortage of Claude Code tips floating around, but the ones worth actually building your workflow around are the ones the team relies on themselves. While the team shared a few more tips, these were the ones that stuck with me!