AI-Powered Prototyping: One Developer's Journey Building a Custom Task Switcher with Claude

A developer, irked by the performance of the Plasma desktop environment's task switcher on X11, embarked on a unique project: building a custom alternative with the help of AI. The default task switcher, particularly in 'Gallery' view, sometimes took up to a second to open, a delay deemed unacceptable for a frequently used feature. Enter FastTab, a custom task switcher built using Zig and OpenGL, designed for instant responsiveness as a daemon process.

The developer, acknowledging this as a potentially 'audience of one' problem—catering to users still on X11 who value performance with many open windows—initially considered accepting the sluggishness. However, the accessibility of AI tools shifted the equation. Leveraging Claude, despite having no prior experience with Zig or X11 internals, a working prototype was created within days, followed by iterations to refine the final product. This highlights a pivotal shift: AI tools enable developers to create solutions they might not have otherwise attempted.

The process began with a detailed conversation with Claude, exploring the problem space and gathering potential solutions. The collaboration then progressed to crafting a detailed specification for FastTab, outlining its functionality and appearance. This specification was iteratively refined until the developer was fully satisfied. Crucially, the author emphasizes the importance of well-structured specifications, favoring pseudocode and visual diagrams (Mermaid diagrams) over extensive code snippets to maintain a focus on architecture and design.

To mitigate risks associated with granting an AI agent access to their filesystem, the developer employed Docker containers. A customized version of 'contai' provided a secure, sandboxed environment for the AI to operate within, preventing accidental data loss from potentially destructive commands. The LLM was informed of its containerized environment, lacking X11 display access or package installation capabilities, to avoid wasted computation on impossible tasks. The author found the containerized approach so effective that it has become their standard practice for all new development projects.

The developer initially experimented with OpenCode and 'oh-my-opencode,' a multi-agent system. However, the multi-agent approach proved less efficient than a single agent, consuming more tokens without delivering significantly better results. Further, changes in Anthropic's terms of service necessitated a return to Claude Code. Token limits remained a challenge, particularly when working with Zig, which the author suspects is more resource-intensive for LLMs due to its low-level nature and limited training data. Both Opus 4.5 and Gemini 3 were used to overcome these limitations.

Despite the AI's assistance, the developer emphasized that AI-assisted coding is not a replacement for coding expertise. While the AI could achieve approximately 80% of the desired outcome, the remaining 20% required human intervention and refinement. This included refactoring code for maintainability, identifying and correcting bugs, and applying domain-specific knowledge, such as utilizing SIMD instructions for performance optimization. Ultimately, the project underscored that AI serves as a powerful tool to accelerate development, especially for personal projects where the goal is to 'scratch an itch' and create something useful quickly.