An Apple developer revealed that they completed 95% of the development work with Claude, and the developed app has been launched on the app store.

An Apple developer reveals using AI to develop applications, with 95% "Claude content"!

Here's what happened. A new native macOS application Context for debugging MCP servers was recently released by an Apple developer —

An application almost entirely built by Claude Code.

The author, indragiek, has been developing software for Mac since 2008.

This time, his goal was to use Apple's SwiftUI framework to create a handy and practical developer tool for the macOS platform.

Different from the past, Claude Code took on 95% of the workload for the Context project. Indragiek claimed:

In this project with 20,000 lines of code, I estimate that I personally wrote less than 1,000 lines of code.

The "engineer" Claude is really getting good. It can work for Apple (doge).

Putting the jokes aside, let's "learn" how this developer uses Claude.

An Apple developer teaches you how to "tame" Claude

As an experienced engineer, Indragie, like many of his peers, has a list of "abandoned projects".

Although he can build project prototypes, the last 20% of the delivery work often consumes a great deal of time and energy, leading to project suspension.

So, he hasn't successfully released any side projects in the past six years.

In February this year, he started trying to use Claude Code to assist with the project, but in the end, Claude almost did all the work for him.

Claude uses the "Agent loop" as the core of development. It only requires inputting prompts through a simple text box, which directly "replaces" the traditional IDE, a branch of VS Code.

In the actual development process, Claude can locate and read the existing source code in the project, understand the code style and design patterns, read the provided additional documentation, generate test validations, compile the program and run tests, and perform iterative repairs based on compilation and test failures.

According to Indragie's feedback, when it comes to using Swift and SwiftUI, Claude performs better in versions before Swift 5.5, especially in SwiftUI.

It can generate accurate but perhaps not very aesthetically pleasing UI code, but the aesthetic issue can be improved through iteration.

Just like Indragie mentioned, simply input in the text box: Make it more beautiful.

Then you'll get an effect like this:

Here comes the key point!

When it comes to major changes like Swift Concurrency and the choice between old and new APIs, Claude sometimes "can't quite grasp it".

So, Indragie created a CLAUDE.md file containing basic instructions on using modern APIs, which can help Claude avoid common "traps".

Indragie emphasized that "context engineering" is very important. He adopted the method of "presetting Agents" to maximize Claude's efficiency.

He found that although the model has a context window of 200k tokens, the model's performance will decline as the use of the context window increases, and the "compression" mechanism may cause important details to be lost.

For this reason, he adopted the method of "presetting Agents", allowing the Agents to pre - read additional context (such as the CLAUDE.md file, specific documents, or source code) to improve the output quality.

Claude also requires detailed requirement specifications.

Any input method such as voice or typing can be used, but Indragie said he prefers typing.

The key is to turn on the extended thinking mode first!

To prevent Claude from blindly jumping into the implementation phase and resulting in poor - quality results, Indragie will ask Claude to first use other extended thinking modes and "make a plan".

By using keywords such as "think" < "think hard" < "think harder" < "ultrathink", different levels of extended thinking of Claude can be activated. Among them, "ultrathink" consumes the most tokens but can produce the best results.

Claude can independently drive the feedback loop, enabling it to make changes, conduct tests, and collect the context of failure reasons.

So, Indragie suggests setting up an effective feedback loop — build, test, fix bugs, and improve the user experience.

He used XcodeBuildMCP to simplify the problems of building and running the application. However, for bugs or UX issues that require user interaction to trigger, manual provision of logs or screenshots is still necessary.

In addition to writing code, Indragie also found that as a general - purpose model, Claude Code can do more than just coding tasks. It also includes editing copywriting, planning functions, etc.

He believes that one of the most useful features is generating realistic mock data, which greatly accelerates the development of UI prototypes and function verification, especially when there is no real data.

Indragie also found that building high - quality automation is almost free.

He asked Claude to write a 2000 - line Python release script, which can check the environment, generate update logs, generate Sparkle appcast (an XML file describing the update information of macOS applications), publish to GitHub, and upload debugging information.

After the script was completed, he used a simple one - line prompt to beautify the CLI output, and finally got this effect:

After this teaching experience, Indragie realized that he mainly used only two tools: Claude Code and GitHub Desktop.

So, he believes that future IDEs will be very different. The source code editor may no longer be the core of the IDE. They will focus on helping developers preset the context of Agents and set up the feedback loop that is crucial for Agents.

After successfully releasing the Context application, Indragie said excitedly:

For me, the most exciting thing in this process is not the application I built, but that I can now satisfy my programming desire again and release excellent side projects. It's like I have an extra 5 hours every day, and the only cost I pay is $200 per month.

One More Thing

According to the data released by Claude Code, since its launch in February this year, it has been used by 115,000 developers, and it has processed 195 million lines of code in a single week.

Assuming Claude code is a junior engineer, these data mean that its annual income can reach $130 million.

The wind of a $100 - million - dollar annual salary has finally reached Claude!

Project address: https://github.com/indragiek/Context

Reference links:

[1]https://www.indragie.com/blog/i-shipped-a-macos-app-built-entirely-by-claude-code

[2]https://news.ycombinator.com/item?id=44481286

[3]https://x.com/deedydas/status/1941683553361854710

This article is from the WeChat official account "QbitAI". Author: Wen Le. Republished by 36Kr with authorization.