Back
Technology

AI Tools Demonstrate Capabilities and Challenges Across Software Development Projects

Generated on: Last updated: 7 sources
View source

Integrating AI into Software Development Workflows

The integration of Artificial Intelligence into software development workflows is being explored across various projects, from application porting to new feature development and local coding solutions. Developers are leveraging AI tools like Claude Code, Xcode's integrated AI, Goose, and Qwen3-coder to accelerate development. Reported benefits include significant efficiency gains, but challenges related to tool management, performance, and iterative refinement are also common.

AI-Assisted Application Development

Mac App Porting with Claude Code

One developer utilized Claude Code to port an existing iPhone application, designed for managing 3D printing filament spools, to a Mac platform. The original iPhone app, built with Swift and SwiftUI in Xcode, tracks approximately 120 spools using NFC tags and camera functions for images. The objective for the Mac version was to enable desktop monitoring of material availability and integrate with print job initiation.

The porting process, despite Swift's UI crossover capabilities, required substantial adaptation due to Mac-specific limitations such as the absence of a camera and NFC scanning. The Mac application was designed to display existing NFC and photo data but not actively manage these inputs. Instant synchronization with the iPhone app via iCloud was a core requirement, facilitated by the existing iCloud architecture. The developer initiated the process by branching the Git repository and instructing Claude Code to develop the Mac app, excluding NFC and camera features but incorporating iCloud.

The initial AI-generated Mac application became functional within ten minutes.

However, it exhibited user interface design deficiencies and performance issues, including slow loading of the spool list. Development proceeded through iterative cycles of building, identifying errors and warnings from Xcode, and providing corrective feedback to Claude Code. Challenges included ensuring UI elements scaled correctly with resizable Mac windows and adapting from the iPhone's screen-to-screen transitions to a three-pane Mac interface.

An issue where the Mac's three-pane display implementation inadvertently affected the iPhone app was resolved by specific instructions to Claude Code to create separate interface screens for each platform. Performance issues related to slow loading of full-resolution photos in the spool list were addressed by implementing a background thumbnail caching system, which generates and stores 128-pixel photo versions.

The developer described the experience as primarily "managing" Claude Code, akin to overseeing a remote human programmer, involving assignments, inquiries, error identification, and extensive testing. The project resulted in a functional Mac application developed over approximately 8 hours of active work spanning one month, which the developer reported as significantly faster than manual coding. Recommendations included slow, incremental feature addition, continuous refinement, and rigorous testing due to the AI's potential to omit requested features.

iOS Sewing Pattern Manager with Xcode 26.3 AI

Another developer reported using Apple's Xcode 26.3 AI for the creation of a new iOS application designed to manage sewing patterns. This application incorporated features such as photography, NFC tag tracking for physical items, and machine learning for image processing. The developer noted that earlier versions, such as Xcode 26.1 AI, had been less effective for complex development tasks.

The project involved migrating an existing codebase and integrating new functionalities. The developer reported the addition of 52,947 new lines of code and the deletion of 10,626 lines across 689 files in less than two days of partial work. This work included integrating Apple's latest AI/ML libraries for image analysis, enabling tasks such as identifying pattern envelopes, straightening images, cropping, and extracting vendor names and pattern numbers via optical character recognition (OCR). The AI was also trained to differentiate pattern numbers from barcodes. The developer estimated that this work would have taken a lone individual 4 to 6 months without AI assistance.

Challenges encountered during this development included issues with Xcode 26.3's integrated AI assistant, specifically its use of multiple background agents, which reportedly led to system hangs, a lack of visibility into ongoing changes, and excessive consumption of token allocation. These issues resulted in work stoppages.

A resolution was implemented by instructing the AI: "Do NOT use background agents or background tasks. Do NOT split into multiple agents. Update me regularly on each step. Do NOT run steps that take more than a minute or two without having an update heartbeat." This instruction reportedly improved productivity and stability.

Xcode's AI assistant offers features such as image pasting, automated code builds and error fixing, and integration with voice dictation tools. Apple is expected to release the full version of Xcode 26.3 on the Mac App Store.

Local AI Coding Solutions

Goose and Qwen3-coder Setup and Testing

A guide detailed the setup and initial testing of a free, local AI coding agent utilizing Goose (an agent framework) and Qwen3-coder (a coding-centric large language model). This setup requires a powerful local machine.

The installation process involved three primary tools:

  • Ollama: An LLM server. Users are instructed to download and install the Ollama application, then select and download the Qwen3-coder:30b model (approximately 17GB) from within Ollama's interface. Ollama's settings require enabling "Expose Ollama to the network" and configuring the context length. Ollama must remain active when Goose is in use.
  • Goose: An agent framework. After installing Goose, users navigate to "Other Providers" and "Go to Provider Settings" to configure Ollama as a provider, selecting qwen3-coder:30b as the model.

Initial testing involved prompting Goose/Qwen3-coder to develop a simple WordPress plugin. The task required five attempts to be completed correctly, with the first three attempts failing to produce a working plugin or follow directions completely. While other free chatbots reportedly complete similar coding tasks in fewer attempts, the agentic nature of Goose allows for iterative corrections by interacting directly with the source code.

Performance varied based on hardware. An M1 Mac with 16GB of RAM reportedly experienced significant performance issues. Conversely, an M4 Max Mac Studio with 128GB of RAM reported good overall performance, with prompt turnaround times comparable to cloud-based or hybrid AI coding solutions for simple tasks. Further analysis with larger projects is suggested to assess if this local solution can serve as an alternative to paid services.

Common Observations in AI-Assisted Development

Across different AI tools and projects, several common themes emerge:

  • Productivity Gains: Developers reported significant increases in development speed, with one project estimating a reduction from months to days of partial work.
  • Management Role: The use of AI tools often shifts the developer's role towards managing the AI, providing instructions, identifying errors, and refining outputs, similar to overseeing a human programmer.
  • Iterative Refinement and Testing: AI-generated code frequently requires multiple iterations, corrections, and rigorous testing to achieve desired functionality and performance. Tools may omit requested features or produce suboptimal designs initially.
  • Hardware Requirements: For local AI solutions, substantial computing power, particularly RAM, is a factor for effective performance.
  • Specific Challenges: Issues observed include difficulties with UI adaptation across platforms, performance bottlenecks due to unoptimized data handling, and complexities in managing multiple AI agents or tasks. Developers have implemented specific instructions to mitigate agent-related problems.

The experiences suggest that while AI coding tools offer potential for accelerated development, they necessitate active management, iterative processes, and a foundational understanding of coding principles to guide the AI effectively.