This article captures the experience of individual developers building "AI factories" using AI coding tools (especially Claude Code), along with candid opinions from various developers presented chronologically. It features a deep discussion about how useful AI actually is, how far automation can go, and what the limitations and practical challenges are.
1. The "Aha Moment" with AI and First Experiences
The article begins with the story of a developer who experienced an "aha!" moment while using AI coding tools like Claude Code. This developer gave Claude Code the danger flag (--dangerously-skip-permissions) and assigned it a complex problem, then was amazed to watch AI leverage multiple tools on its own to solve it.
"The 'aha' moment is when you give Claude a hard problem and watch it spend several minutes running different tools by itself to solve it."
This developer demonstrated AI compiling, running, and debugging a 486 assembly Mandelbrot fractal generator in Docker, and the results were satisfactory.
2. The "Trivial Example" Controversy and AI's Limitations
Several developers then pointed out that such examples are actually too simple (trivial) problems. Assembly and Docker environments are already well-represented in AI training data, and complex real-world problems are still beyond AI's capabilities.
"The hardest problem in computer science might be 'creating a non-trivial LLM demo.'"
Another developer noted that AI excels only at tasks with clear success criteria that are easily verifiable.
- In reality, 75-99% of a software engineer's work consists of such simple tasks.
- But designing and verifying complex systems remains a human responsibility.
"If AI can do what would take weeks for under $10 in one hour, that's quite valuable."
3. The Boundary Between "Trivial" and "Non-Trivial"
This leads to a discussion about the definition of "non-trivial."
- A non-trivial example means a complex system that requires specialists from multiple fields working over several months.
- Even when AI produces results for such examples, humans must personally verify and redesign, so it actually takes a long time.
"A non-trivial example is not something where you look at the output and immediately judge 'AI did well.' It's a task where a human has to dig in, understand the complexity, and verify."
For this reason, AI's true limitations become apparent when solving completely new problems not found in training data.
4. AI in Real-World Production and Its Limits
Many developers share their experiences of how useful AI actually is in practice.
- Simple code generation, repetitive tasks, and documentation can be handled quickly and cheaply by AI.
- But complex systems, UI/UX, architecture design, and maintaining consistency still require direct human involvement.
"Even when AI generates code well, I always review the changes myself and optimize. If you just leave it, bugs appear and customers suffer."
Additionally, running multiple AI agents simultaneously causes each to solve problems in completely different ways, breaking consistency.
"Each agent approaches with completely different ideas, so the user experience becomes inconsistent. You end up having to create thousands of lines of instruction files, which actually makes things more complex."
5. Real AI Coding Workflows and Concerns
Several developers introduce their own AI coding workflows.
- Approaches where multiple agents converse to solve problems,
- Approaches that create a plan document first, clarify through questions, then generate code,
- Approaches that delegate small tasks to AI while managing important parts directly, among various attempts.
"I always have it ask questions first, create a
spec.mddocument, and then write code. This is much more efficient."
However, these workflows also have clear limitations in complex large-scale projects.
"It really helps with small tasks, but for more complex work, I had to intervene directly most of the time. Ultimately, the final output is my responsibility."
6. Diverse Perspectives on AI's Value and Future
Opinions on AI's value are sharply divided.
- The position that "AI is already sufficiently valuable" and
- The skepticism that "it's still an expensive hobby, limited to repetitive task automation" coexist.
"AI handling simple tasks quickly is truly powerful. But in the end, it's just work that millions of tired people could also do."
"I can't feel attached to AI-generated code. I end up working on important projects myself and just abandoning the rest."
There's also the opinion that AI-generated code is ultimately disposable, and the real assets are the plans, prompts, and agent combinations.
"Code is disposable. The real assets are plans, instructions, and agents."
7. Real Costs of AI Coding and Subscription Issues
Practical concerns about subscription plans and costs of AI coding tools also arise.
- Claude Max and other premium subscriptions can handle heavy usage.
- But Pro plans quickly hit their limits, frustrating some users.
- Running multiple sessions in parallel drains the limit even faster -- a tip that's shared.
"Running multiple sessions on the Pro plan hits the limit in less than an hour. Running one at a time is fine for 3-4 hours."
8. The Future of AI Coding and the Developer's Role
Finally, the point is emphasized that even as AI advances, the developer's role remains important.
- A world where AI automates everything may come, but
- For now, developers must personally handle design, verification, consistency, and architecture management.
"A senior developer always has to play the senior role. Whether it's AI or people, someone needs to set the direction and build the structure."
Concerns about the quality, consistency, and maintainability of AI-generated code also persist.
"AI-generated code sometimes falls into strange patterns, and I often end up having to tear it apart and redo it."
Key Concepts and Summary
- Trivial vs Non-Trivial: The boundary between what AI does well and what it doesn't
- AI Coding Workflow: The cycle of planning-questioning-code generation-verification
- Disposable code, accumulating prompts and plans
- The reality of costs and subscriptions
- Difficulties with consistency, quality, and verification
- Changing developer roles and their limits
- The debate over AI's real value and future
Conclusion
Through this article, it becomes clear that AI coding tools significantly boost developer productivity, but their limitations in complex problems and large-scale systems remain clear. The vision of an "AI factory" is certainly attractive, but human roles in planning, verification, and maintaining consistency remain crucial. Many developers agree that continuing to experiment with how to work alongside AI and evolving individual workflows is the challenge ahead.
"Programming has changed forever. It's best to embrace this change quickly."