AI-Assisted Development: Building with Jaspr and Claude

Building our website with Jaspr offered a unique perspective on AI-assisted development with emerging frameworks. While I've used AI tools extensively in other projects, working with Jaspr — a relatively new framework — presented different challenges and insights. Here's what I learned about AI-assisted development when the framework itself is still evolving.

The SDLC Reality: Humans at Both Ends

When we think of the Software Development Life Cycle (SDLC) as a loop from concept to release, it's important to recognize that — at least for now — both ends of this cycle remain fundamentally human-dependent.

The concept phase requires understanding business needs, user problems, and strategic direction. The release phase demands judgment calls about quality, user experience, and timing. AI can assist, but it cannot own these decisions. The middle phases of planning, implementation, and verification? That's where things get interesting.

Planning: The 80/20 of Development

After settling on our concept, I experimented with having AI create implementation plans. This turned out to be one of the most valuable parts of the process. The quality of the planning phase directly correlates with time saved during implementation — garbage in, garbage out.

I used Claude Sonnet 4.5 and Claude Opus 4.5 for planning, providing detailed context about our goals, constraints, and architecture. The more time I spent refining the plan with AI, asking clarifying questions, and challenging assumptions, the smoother the implementation became.

// Example planning prompt
"We need to implement an auto-scrolling tech showcase.
Requirements:
- Edge-based activation (20% zones)
- Variable speed based on cursor position
- Seamless infinite loop
- Works on mobile with touch"

The AI's responses weren't just code suggestions — they were architectural discussions about UX patterns, performance considerations, and implementation strategies. This collaborative planning saved hours of refactoring later.

Implementation: When AI Meets Novel Tech

In my experience with established frameworks like Flutter, AI tools handle most implementation work smoothly. But Jaspr presented a different challenge as relatively new technology. Two specific issues emerged:

• Limited training data: AI models haven't been extensively trained on Jaspr patterns
• Hallucination risk: AI can't say "I don't know," so it sometimes confidently suggests incorrect Jaspr APIs

This meant I had to be more hands-on than usual. I found myself reading Jaspr documentation, understanding the framework's mental model, and writing initial scaffolding code by hand — almost like the old days!

// Had to manually write initial patterns
@client
class TechScroller extends StatefulComponent {
  // After establishing this pattern,
  // AI caught on quickly for similar components
}

But here's the fascinating part: once I established the main skeleton and patterns, AI picked up on them remarkably fast. It's like teaching someone through examples — after seeing a few @client components and state management patterns, the AI started generating consistent, correct Jaspr code.

Context7: Up-to-Date Documentation for AI

One tool that significantly improved AI assistance was Context7, an MCP (Model Context Protocol) server that provides up-to-date documentation and code to LLMs and AI code editors like Cursor and Claude Code.

The challenge with new frameworks like Jaspr is that AI models are trained on data with a cutoff date. Context7 solves this by fetching the latest documentation, examples, and API references directly into the AI's context. Instead of the AI hallucinating outdated or incorrect Jaspr APIs, it can reference current documentation.

This was crucial for our project. When I asked about implementing client-side hydration or state management, Context7 ensured the AI was working with the latest Jaspr patterns and best practices, not outdated information from its training data.

Tools of the Trade

For this project, we used a combination of AI tools, each bringing different strengths:

Claude Code

Our primary development environment. Claude Code excels at understanding context across multiple files, making architectural decisions, and maintaining consistency. We used both Claude Sonnet 4.5 (for fast iterations and standard implementations) and Claude Opus 4.5 (for complex architectural decisions and novel problems).

GitHub Copilot

Copilot served as a valuable companion for coding questions, refactoring suggestions, and exploring alternative implementations. It's particularly helpful for getting quick answers without switching context.

Verification: The Human Safety Net

In a typical modern web project using React or Vue, AI can handle verification almost 50/50 with humans — running tests, checking console errors, validating against requirements. But with Jaspr's dual-environment model (server and client), I ended up taking on about 80% of the verification work.

Why? Because errors in Jaspr can be subtle:

• Code that works on client but crashes on server during SSG
• Import issues between dart:html and universal_web
• Hydration mismatches between pre-rendered and client-rendered output
• Browser API usage without proper kIsWeb guards

These aren't bugs that AI necessarily catches in its first pass. They require understanding Jaspr's architecture and knowing what to look for. As the ecosystem matures and AI models get trained on more Jaspr code, I expect this ratio to shift back toward 50/50.

Lessons Learned

1. Plan More, Code Less

The best ROI came from spending extra time in the planning phase. A well-structured plan with AI guidance meant less backtracking during implementation.

2. Establish Patterns Early

With new frameworks, manually write 2-3 exemplar components. Once AI sees the pattern, it maintains consistency remarkably well.

3. Know When to Read Docs

For bleeding-edge tech, AI might not have the answers. Being able to quickly read official documentation and translate it into working code is still a crucial skill.

4. Use Multiple AI Tools

Different AI tools excel at different tasks. Claude Code for architecture and complex problem-solving, Copilot for quick coding assistance, Context7 for up-to-date documentation access.

5. Verify Religiously

Especially with new frameworks, don't trust AI output blindly. Test in both server and client environments, check for edge cases, and validate against documentation.

Reflections on AI-Assisted Development

This project highlighted how AI-assisted development changes when working with emerging frameworks. With established technologies like Flutter, AI tools confidently handle most implementation. With Jaspr, the collaboration required more careful oversight and verification.

The sweet spot I found was using AI for planning discussions, pattern consistency, and implementation acceleration, while relying on human judgment for architectural decisions, framework-specific nuances, and verification.

What surprised me most was how the process changed my development workflow:

• Spending more time on planning and architecture upfront
• Using AI as a discussion partner for design decisions
• Being more deliberate about establishing code patterns
• Maintaining a healthy skepticism while verifying outputs

Building our website with Jaspr and AI assistance was a learning experience. AI tools like Claude Code and Copilot significantly accelerated certain phases of development, particularly once initial patterns were established. But the project also reinforced the continued importance of understanding the underlying technology, reading documentation, and applying critical judgment throughout the development process.