Understanding AI

Understanding AI

Share this post

Understanding AI
Understanding AI
Claude-powered coding tools are poised to transform programming
Copy link
Facebook
Email
Notes
More

Claude-powered coding tools are poised to transform programming

A deep dive into the powerful tools that were made possible by Claude 3.5 Sonnet.

Timothy B. Lee's avatar
Timothy B. Lee
May 29, 2025
∙ Paid
51

Share this post

Understanding AI
Understanding AI
Claude-powered coding tools are poised to transform programming
Copy link
Facebook
Email
Notes
More
4
1
Share

An underrated AI story over the last year has been Anthropic’s success in the market for coding tools.

“We believe coding is extremely important,” said Anthropic engineer Sholto Douglas in an interview last week. “We care a lot about coding. We care a lot about measuring progress on coding. We think it’s the most important leading indicator of model capabilities.”

This focus has paid off. The company’s models have excelled at software engineering since last June’s release of Claude 3.5 Sonnet. Over the last year, a number of Claude-powered coding tools—including Cursor, Windsurf, Bolt.new, and Lovable—have enjoyed explosive growth. In February, Anthropic released a coding assistant called Claude Code that has become popular among programmers.

In media interviews, Anthropic employees have touted the extreme efficiency gains Claude has enabled for its own programmers.

“For me, it’s probably 2x my productivity,” said Anthropic engineer Boris Cherny in a recent podcast episode. “I think there’s some engineers at Anthropic where it’s probably 10x their productivity. And then there are some people that haven’t really figured out how to use it yet.”

Cat Wu, an Anthropic product manager, chimed in with an example: “Sometimes we're in meetings together and sales or compliance or someone is like, hey, like, we really need X feature. And then Boris will ask a few questions to understand the specs. And then like 10 minutes later, he's like, all right, it's built. I'm going to merge it later. Anything else?”

Anthropic’s success in the coding market has gotten the attention of both OpenAI and Google:

  • In early May, OpenAI announced it was acquiring Windsurf, an AI-powered code editing tool that had been powered by Anthropic models.

  • The next week, OpenAI announced Codex, a coding agent designed to compete with Anthropic’s Claude Code.

  • Last week Google announced its own coding agent called Jules.

I suspect one of Anthropic’s major goals for Claude 4, which was released last week, was to maintain its lead in this market. It seems to be helping. Days after the release of Claude 4, the CEO of vibe coding tool Lovable wrote that “Claude 4 just erased most of Lovable's errors.” He posted a chart showing a dramatic drop in syntax errors after Lovable upgraded to Claude 4.

In recent weeks, I’ve talked to a number of software developers and product managers about how AI-powered tools have changed their work. Based on these conversations, I think we’re on the verge of dramatic changes in the way people create software.

In this piece, I’ll survey the new software development tools that have gotten traction in the last year. I’ll start with “vibe coding” tools designed to enable programming novices to build full-featured apps. Then I’ll discuss tools designed for experienced programmers. As we’ll see, the leading tools in both categories owe their success to Claude.

The vibe coding revolution

Last week I talked to Anthony Jantzi, a product manager at Gloo, a startup that creates software for churches and other Christian organizations.

“In the olden days we would use Figma to build interactive prototypes where you could click around and it looked like a web app,” Jantzi told me. But he recently started using a vibe coding platform called Bolt.new for prototyping and it dramatically changed how he did his job.

Those old Figma mockups looked like real websites, but a lot of features didn’t actually work. For example, Gloo’s website includes a chatbot, which is beyond the abilities of a Figma mockup. So although Jantzi could solicit user input about the layout of a new feature, he told me it was “impossible to get any kind of good feedback” about its functionality.

Jantzi created a fully functional clone of Gloo’s website using Bolt with just a few weeks of work.

“I basically have built a prototype version of our app in Bolt that I can basically make whatever changes I want to, and put it in front of a potential user and see how they use it,” he told me.

Bolt lets Jantzi add new features with plain English prompts. It takes a fraction of the time it would take with conventional programming tools. But Jantzi said he wouldn’t use Bolt (or other vibe coding tools) to build a shipping software product.

“It's not getting to the level of robustness of an actual app,” he told me. “If I put it out for someone to use with any kind of volume of users it would fall over.”

So Jantzi still needs help from traditional engineers to put new features into production. But testing a feature first in his Bolt sandbox lets Jantzi use their time more efficiently.

“I'm not having my engineers waste time on things users won't want,” he told me.

Jantzi’s story isn’t unusual, according to Eric Simons, CEO of the company behind Bolt.new.

Keep reading with a 7-day free trial

Subscribe to Understanding AI to keep reading this post and get 7 days of free access to the full post archives.

Already a paid subscriber? Sign in
© 2025 Timothy B Lee
Privacy ∙ Terms ∙ Collection notice
Start writingGet the app
Substack is the home for great culture

Share

Copy link
Facebook
Email
Notes
More