🚀 AI-Assisted Software Engineering: My Experience
I’ve been experimenting with using AI tools like Cursor and ChatGPT to help design and refine product features, database schemas, and documentation. Here’s what I’ve learned so far, and why I think the right approach is what I call AI-assisted software engineering.
📝 Brain Dump to Database
I started by dumping out features and descriptions of my product — adding more detail and nuance each time I revisited it. Then I used TypeCPT to generate potential tables and prompts. Feeding that into Cursor, I asked it to design a database, and it actually produced a schema. I then asked it to document the features of each table. Suddenly, I had a rich starter description with tables I could use for my MVP.
This sped up the process tremendously, but also showed me the dangers of following AI output blindly. You still have to read, interpret, and refine what comes back.
⚖️ Traditional Approach vs AI
Normally, a team of developers, analysts, and business leaders spend weeks mapping out business processes, capturing data requirements, designing interfaces, and translating those into a database, feature set, and roadmap. They then test and iterate along the way. This is time-consuming because features are interdependent, leading to complexity and waterfall effects.
AI changes that by giving instant answers and draft structures — but not perfect ones. I’ve seen about 75–85% accuracy. It will never be 100% because your input isn’t fully realized either. If you had a fully detailed spec with tasks and data models, you wouldn’t need the AI for that part — you’d just use it to generate the UI and file structure.
💡 The Real Promise (and the Catch)
People want AI to do everything: the data model, the PRD, and even the roadmap. But the truth is, to get great output you need great input. That means the user has to become more skilled at refining specifications with more detail and granularity. Yes, that slows things down. But it’s worth it — because then you get exactly what you want.
🛠️ My Shortcut
Here’s the shortcut I recommend: take a competitor’s feature set, clone it, describe it, and build your PRD from that. Then add your innovations, improvements, and competitive edges — backed by research, intuition, and industry knowledge. This gives you a strong foundation and ensures you’re building something better, not just copying blindly.
📌 Conclusion
There’s no true shortcut in software development — only smarter processes. AI helps accelerate and augment the work, but it doesn’t replace the need for clear, detailed thinking. That’s why I call it AI-assisted software engineering, not “vibe coding.” It’s not just a marketing buzzword. It’s a real shift in how we can design and build software.
The Rise of Spec-Driven Development in the AI-First Era
There’s a major shift underway in how products are built in the AI-first era. For years, many developers have relied on "vibe coding" — building in the moment without a clear structure. But as the industry matures, it’s clear that this approach doesn’t scale. Spec-driven development is emerging as the professional, reliable way forward: a structured workflow that takes ideas from concept to customer consistently. In this post, I’ll walk through why this matters, how the industry is catching on, and what challenges still remain as we refine this new way of building.
Why Spec-Driven Development Matters
Spec-driven development represents a shift away from unpredictability toward a disciplined process. As a business owner and product designer, I’ve seen firsthand how it not only changes how we code but how we build products altogether. Instead of rushing straight into writing specs, the process begins with alignment: ideation, trade-off discussions, and strategic planning. This ensures that when the spec is finally written, it reflects both the product vision and customer needs with clarity.
The Industry Is Catching On
Big toolmakers are now converging on the same realization. GitHub launched their Spec Kit, Amazon released Kirao built around planning with specs, Cursor added task tracking, and Cloud Code has planning mode. Across the board, the industry is beginning to recognize that specs are essential. But while momentum is building, the tooling and processes behind spec-driven development are still unsettled. And that’s a good thing — we’re collectively figuring out what it means to build in this new AI-first era.
Three Problems Holding Back AI Adoption
From applying spec-driven development in my own products with Agent OS, I’ve identified three fundamental problems most tools don’t address:
- Lack of alignment process — Tools often assume you know what you want to build, skipping the crucial phase of strategy and trade-offs before the spec is written.
- Missing context about standards — Specs need to include coding standards, patterns, and existing libraries so AI agents can build correctly the first time.
- Rigid workflows — Teams build differently. Tools that impose one rigid process prevent adoption and reduce flexibility across projects.
How Agent OS Tackles These Challenges
Agent OS introduces a more adaptive process. It begins with clarifying questions and strategic planning before specs are written, incorporates coding and design standards, and uses sub-agents to handle specific parts of implementation. This flexibility allows each feature to follow its own path — whether API-first, UI-first, or something entirely unique — without breaking the workflow.
The Future of AI-First Development
The real power of spec-driven development isn’t just efficiency — it’s about preserving our craft. While AI agents handle repetitive tasks, we as builders bring value through design, judgment, and vision. By embedding standards, context, and flexibility into the process, we can achieve flow states again and unlock breakthroughs in product development.
Comments
Post a Comment