Experiment Begin: The No-Code-Touch Constraint Play
I’m currently rebuilding my tool suite in conjunction with the recent site renewal.
These were originally self-made tools, but this time I decided to deliberately reconstruct them using an AI-driven approach. The reason is simple and clear: “The UI will be more refined than what I prepared myself.”
I used Claude Code as the generative AI. And here’s the rule I imposed on myself:
Read the generated code, but never modify it myself. Complete everything through prompt instructions only.
It’s essentially constraint play. How far can we go?
First Impression: Beyond Expected Refinement
First, regarding UI - it exceeded expectations. It definitely gets refined.
Compared to the gritty utility-style UI I created myself, the interfaces generated by Claude Code are clearly modern and beautiful. Responsive design is naturally incorporated, and accessibility considerations are evident.
“Wow, this is certainly impressive.”
The initial amazement was definitely there.
Gradually Emerging “Claude-ness”
However, the more tools I created, the more déjà vu began to emerge.
They all start to have similar “shells.” And for some reason, it seems to favor blue and purple gradient-like atmospheres. Well, they’re pretty colors, so that’s fine.
This area really shows the “made with Claude Code” feel, so there’s still room for improvement in visual approaches.
It’s like the same designer reusing templates.
Single-Function Tools: “Almost One-Shot” Power
What gets created generally works well, and explanations are very accurate. Honestly, as an assistant, it’s extremely competent.
For single-function compact tools, it probably delivers about 90% of expected value in almost one shot. Really flawless.
“Most things are probably fine with just this.”
That’s how it makes me feel. If something satisfactory gets completed, that should be OK.
However, When You Step Into the Swamp of Perfectionism…
But when it comes to “making it even better,” “adding features,” or “seasoning it to my taste,” the story completely changes.
To go further from here, you need to steel yourself and face AI anew.
Or rather, it becomes no different from dealing with humans.
The Pro vs Pro Wall: Difficulty of Communication
When things turn out different from what you expected, conveying “I want it like this” to get the desired output is extremely difficult.
If you were an amateur, you couldn’t make such precise good/bad judgments, so it would end with “that’s amazing.”
But when you become knowledgeable in the field or develop strong preferences, that’s no longer the case. The more it becomes pro vs pro, the more you notice the gap, the magnitude of that divide.
The more creative the activity, the more pronounced this becomes.
Maybe It’s the Same as Architects and AI
This might be exactly the same as architecture, painting and sculpture arts, music, literature, theater, or even the culinary world.
When something that could be done by one person suddenly has multiple creators, there comes a moment when it won’t take shape unless you communicate correctly.
- The ability to see the goal
- The organizational skills for that process
AI’s advanced development tests exactly this.
Even if AI were an omniscient god, if your ability to communicate is weak, you won’t get the results you want at all.
Commonalities with Human Management Work
Setting aside specific methodologies, I strongly felt that this is no different from communicating with humans in management work.
- Estimating the other party’s understanding level
- Conveying information at appropriate granularity
- Receiving feedback and making course corrections
- Continuously communicating toward the goal
It’s project management itself.
Beyond the Dimension of “Jobs Being Stolen”
The talk about AI stealing programmers’ jobs is no longer at that dimension.
To manage generative AI and achieve results requires:
- Much higher resolution understanding
- Ability to see ahead
- Ability to clearly imagine goals
- The precision of these determines success or failure
In other words, I concluded that broader knowledge and experience are required.
Combat Data: Production Time and Difficulty
Here’s the actual production time and perceived difficulty. Of course, I mostly adhered to the “never touch code” constraint.
Difficulty ⭐️: The World of Almost One-Shot
- Production time: About 10 minutes
- Instruction count: Almost 1 shot
- Difficulty: ⭐️
- Production time: About 20 minutes
- Instruction count: 2-3 shots
- Difficulty: ⭐️
- Production time: About 30 minutes
- Instruction count: 5-10 shots
- Difficulty: ⭐️
QuickQR Enhanced - QR Code Generation/Reading Tool
- Production time: About 120 minutes
- Instruction count: 15-20 shots
- Difficulty: ⭐️⭐️
This area really progresses smoothly. It’s Claude Code’s forte.
Difficulty ⭐️⭐️⭐️⭐️⭐️: The Beginning of Hell
PhotoFlow - Image Analysis/Editing Tool
- Production time: About 5 days
- Instruction count: 100+ shots
- Difficulty: ⭐️⭐️⭐️⭐️⭐️
This is where I was brought back to reality.
80% Completion and the Identity of the Remaining 20%
Personally, I consider all tools to be about 80% complete. They’ve reached a level where they’re normally usable.
So what’s the remaining 20%?
This is some kind of insufficiency, a part where the AI-driven approach feels too instant in its seasoning.
While they work as tools for casual use, there’s a sense of “one more step” - whether this is technical or related to planning and product specifications, some insufficiency remains.
Exploring the Identity of the “Instant” Feel
I thought about the identity of this “instant feel.”
1. The Trap of Predictability
Claude Code is very competent, but has a certain “predictability.” It tends to take similar approaches to similar problems.
While this is good in terms of stability, it also means less “surprise” and “originality”.
2. Depth of Context
AI derives optimal solutions based on given information, but there are still areas where it doesn’t match humans in tacit knowledge and deep contextual understanding.
Considerations for “the mindset of tool users” and “subtle nuances of usage scenarios” tend to remain superficial.
3. Absence of “Soul”
Even when creating something technically near-perfect, you can’t feel the creator’s “soul”.
The warmth of handcraft, the creator’s personality, a touch of playfulness. These elements inevitably become diluted.
Implications for Future Product Development
What became clear through this experience is the importance of role division between AI and humans.
What AI Excels At
- Technical implementation
- Applying general UI/UX patterns
- Realizing standard functions
- Ensuring code quality
What Humans Should Handle
- Original ideas
- Deep contextual understanding
- Emotional connections with users
- Expressing brand value
PhotoFlow Development: 5-Day Battle Record
Let me reflect in detail on the particularly memorable “PhotoFlow” development.
Day 1: Optimistic Start
“Create an image analysis and editing tool.”
I started with a light heart. Since single-function tools were being created smoothly, I thought something a bit more complex would be fine too.
Day 2: Reality’s Wall
“Hmm, this is different from what I expected…”
The more detailed requirements I conveyed, the more it seemed to drift away from expectations.
Day 3: Prompt Hell
“Not like this, not like that…”
Changing instruction methods, changing angles, showing examples… This was truly the beginning of prompt hell.
Day 4: Finding a Breakthrough
“Ah, trying to convey everything at once is the problem.”
I switched to a strategy of dividing functions and implementing them step by step. This changed the flow.
Day 5: Finding a Compromise
“Let’s call 80% good enough.”
I realized that aiming for perfection might never end.
New Skill Set for AI Collaborative Development
Through this experience, I felt that AI collaborative development requires a new skill set.
1. Prompt Design Ability
- Instructions at appropriate granularity
- Building staged requirements
- Efficient context communication
2. Project Division Ability
- Appropriate function division
- Organizing dependencies
- Clear prioritization
3. Quality Judgment Ability
- Instantly judging the quality of AI output
- Quick decision on correction direction
- Identifying compromise points
4. Patience and Flexibility
- Dealing with situations that don’t go as planned
- Adapting to approach changes
- Mental preparation for long-term battles
Changing Value of “Creating”
In the AI era, I feel the value of “creating” itself is changing.
Traditional Value
- Technical implementation ability
- Efficient coding
- Programs with few bugs
New Value
- Original ideas
- Deep user understanding
- Creating emotional value
- Storytelling ability
To Fill the Remaining 20%
So what’s needed to fill that remaining 20%?
1. Injecting Humanity
We might need a process of adding human “seasoning” to the AI-created base.
- A touch of playfulness
- Thoughtfulness toward users
- Orchestrating unexpected discoveries
2. Adding Narrative
It’s important to design the entire experience of using it as a story, not just the tool itself.
3. Continuous Improvement Culture
We need a culture of observing how things are used and continuously improving, not ending with AI generation.
Future Predictions for AI-Driven Development
Based on this experience, let me predict the future of AI-driven development.
Short-term (1-2 years)
- Single-function tool development becomes completely AI-led
- Prompt engineering skills become essential
- Human role centers on direction
Medium-term (3-5 years)
- Complex systems become developable through AI collaboration
- “AI Management” established as a new profession
- Design personalization technology develops
Long-term (5+ years)
- AI understands human creativity more deeply
- From collaboration to partnership
- New forms of creation emerge
Future Challenges and Prospects
Technical Challenges
- Improving AI creativity
- More natural dialogue interfaces
- Long-term memory and context retention
Human-side Challenges
- Systematizing prompt design skills
- Project management methods suitable for AI collaboration
- Redefining creative value
Conclusion: Toward a Future Dancing with AI
Through the 5-day AI-driven development experiment, I learned many things.
AI is certainly a powerful partner. For simple tasks, it delivers results at speed and quality surpassing humans.
However, the more creative and complex the task becomes, the more the importance of the human role becomes highlighted.
Rather than having jobs stolen by AI, an era of creating new value together with AI has begun.
To do this, we need to acquire new skill sets and explore new ways of working.
The remaining 20% dissatisfaction might actually indicate future possibilities.
That 20% might be the most valuable part that only humans can fill.
The tool suite introduced in this article was actually developed in collaboration with Claude Code. Please explore new relationships with AI yourselves.