Claude Code Built a Laravel App From Upwork: Things I've Learned

Diving into AI-assisted coding can feel like uncharted territory, but as someone who thrives on hands-on experience, I find real-world projects are the ultimate proving ground. "AI Coding Daily" took on an Upwork gig—a musician portal for managing gigs and staffing—to push the boundaries of what's achievable with cloud-based coding. After a concentrated three-hour session, I walked away with a functional portal and some crucial insights into optimizing AI-driven development workflows.

The Musician Portal: A Cloud Code Creation

The project's goal was to create a platform where musicians could register for gigs, and administrators could manage the entire process, from assigning musicians to handling logistics. The result was a portal where musicians can view profiles, past gigs, and accept or reject new opportunities. On the administrative side, a Filament-powered panel allows for managing musicians, instruments, assignments, and even includes audit logs and settings. This wasn't just about churning out code; it was about refining a process.

Phase-by-Phase: A Strategic Breakdown

The key to success was breaking down the initial job description into detailed project phases. Rather than overwhelming Cloud Code with the entire scope, I fed it focused tasks, prompting it to work on specific phases (e.g., "phase 3.1"). This iterative approach, refined through about 20 prompts, led to the final product. It's a testament to the power of structured planning when working with AI.

Context is King: Navigating the AI's Awareness

One of the most significant lessons revolved around context management. As "AI Coding Daily" progressed through the phases, it became clear that the context provided to Cloud Code was critical. Each task involved Cloud Code analyzing the existing codebase, documentation, and relevant information. While the initial settings accounted for about 15% of the context, additional details quickly added up. While there was a concern about exceeding the context window, most phases used between 70% and 80% of the available context. However, in later phases, "AI Coding Daily" started encountering context limitations, which can lead to the AI 'hallucinating' or making inaccurate suggestions.

To avoid this, scope of the task must be kept manageable. For example, generating database-related elements (factories, seeders, models, migrations) in one phase worked well because these components are tightly related. In contrast, a phase like gig management, which involves separate tasks like creating, listing, editing, and deleting gigs, required more careful division into sub-phases. "AI Coding Daily" realized that if Cloud Code works for more than 10 minutes on a single task, it's likely to run out of context. The takeaway? Break down tasks into smaller, reviewable chunks.

The Power of Granular Testing

The final lesson came from a simple yet profound change in the guideline prompt: the inclusion of acceptance criteria, essentially tests, within each phase. Previously, Cloud Code generated some tests, but they weren't granular enough. By explicitly instructing Cloud Code to generate specific tests, the development process became more robust. Although it increased the initial development time due to failing tests requiring correction, the end result was significantly more stable. Running a full test suite revealed 566 tests, including browser tests for mobile viewports, all passing. This emphasis on thorough testing drastically reduced the number of bugs encountered during review, proving that more tests lead to more stable code.

The Road Ahead: Embracing AI with Structure

This experiment underscores the importance of structured project planning and context management when using AI coding tools. By breaking down projects into manageable phases and emphasizing granular testing, we can harness the power of AI to create stable, functional applications. I'm eager to explore more context management and optimization techniques. The key is to find the balance between AI assistance and hands-on oversight, ensuring the final product meets the highest standards of quality and reliability. After all, the magic lies in the build.