AI-Assisted Development: Building Niche Tools Like FastTab for an Audience of One

The Hyper-Specific Problem

FastTab addresses a remarkably niche performance issue: reducing the ~1-second delay when activating KDE Plasma's 'Gallery' task switcher on X11 systems. Developer Andrea Dragotta explains this targets an "audience of one" - users who:

• Remain on X11 (rather than Wayland)
• Prefer visual window previews over text lists
• Demand near-instant response from frequently used UI components

Without AI assistance, Dragotta admits he'd have tolerated the lag rather than build a solution requiring unfamiliar territory: Zig programming, OpenGL rendering, and X11 internals.

The AI Development Workflow

The process followed a structured approach:

1. Conversation-First Specification Instead of diving into code, Dragotta started by discussing architectural options with Claude. Key refinements:

   • Purging premature code snippets from specs
   • Using Mermaid diagrams instead of ASCII art
   • Breaking implementation into milestones

     "Pseudocode is fine... it allows focus on architecture without bogged-down implementation details."

2. Safe Execution Environment To prevent catastrophic rm -rf / mishaps:

   • Used heavily customized contai Docker wrapper
   • Restricted filesystem access to project directory only
   • Explicitly informed LLM about container limitations (no X11 access)

3. Token Management Reality Despite initial excitement about multi-agent systems:

   • Found single Claude/Gemini agents equally effective
   • Zig's low-level nature consumed tokens rapidly
   • Hit regular limits requiring model switching

The 80/20 Rule of AI Coding

While AI generated an initial working prototype:

• First version was a 1700-line monolith with duplicated code
• Required significant refactoring for maintainability
• Critical optimizations (like SIMD instructions) needed human prompting

Dragotta notes: "Asking the right questions requires coding knowledge. You need to know what questions to ask to get the best results." Performance-critical sections like window texture handling ultimately borrowed directly from X11 internals after iterative refinement.

Implications for Side Projects

This case highlights shifting possibilities:

Traditional Constraints	AI-Assisted Approach
Unfamiliar tech stack → prohibitive	Prototype in unknown language in days
Niche use = not worth effort	Viable for "audience of one" solutions
Polish requires excessive time	Functional core enables incremental refinement

"For side projects, I can spitball with Claude about my problem, get a spec, and have a working prototype in a few hours," Dragotta observes. The shortened feedback loop increases completion likelihood for tools serving personal needs.

Practical Takeaways

1. Protection First: Always sandbox AI code execution (containers + git)
2. Specs Over Snippets: Focus LLMs on architecture before implementation
3. Embrace Hybrid Work: AI for boilerplate/brainstorming, humans for critical path logic
4. Token Awareness: Budget for LLM limitations, especially with less common languages

The FastTab source code demonstrates this workflow's output – a Zig daemon using XFixes and OpenGL to render window previews at sub-100ms speeds. While not replacing professional development practices, AI assistance now makes hyper-personalized tooling pragmatically achievable.