[🧑‍💻] Building Fast with AI Is Not the Same as Building Professionally

✨ GPT-5.6 Sol’s Summary 　

I sped up implementation with AI and harnesses but missed coherence across planning, design, and documentation, then confronted my arrogance after seeing an external team’s Figma work and manual.

People really do need humility.

A few days ago, when I saw a business system that the company had entrusted to an external team and that had taken quite a while to build, I honestly thought, “If I had built this, I would have finished it much faster…”

The speed I gained from AI and harnesses

Even though vibe coding with AI has plenty of weak spots, I thought it could be improved enough. If I proceeded as carefully as possible with TDD and built good harnesses to prevent recurring issues, that should work. Looking at Big Tech companies that are serious about AI also made me think the share of developers who code directly might keep shrinking.

To be clear, that opinion has not changed at all. If AI is used well, tests and harnesses are laid down properly, and recurrence-prevention structures are built carefully, I still think development speed and quality can change dramatically.

But I did feel that I needed to reflect on the arrogance behind “I would have finished it much faster…”

Recently, I have often developed like this. I go back and forth with agents, keep improving the harness engineering structure, prevent various issues from recurring, add new features, take feedback, and fix things again. For repeated work, I create worker/coordinator agents and run them.

The coherence my development left out

That routine gave me fast development speed and quick changes of direction, and it also did a decent job of preventing similar issues from recurring. But it was weak at tying the whole product together under one planning and design philosophy. I could push implementation quickly, but the consistency across design, functionality, and product planning was very low.

Still, I do not see this as an inherent limit of vibe coding. It is closer to the fact that I have not yet sufficiently embedded standards for planning, design, and documentation into my harness engineering.

Professionalism in Figma and the manual

But when I saw that external team actively using Figma to record different screen structures and leave behind descriptions and flows, something became clear. A professionalism different from my way of charging in without enough structure. The high consistency of their plans and deliverables created a strong sense of trust.

Their professionalism also showed in the GitBook manual. I had only thought, “If I build it well enough, users will figure it out. If the UI/UX is intuitive, that should be enough.” I had never even imagined making a guidebook for internal staff.

Of course, if the UX is good and help buttons are placed exactly where people expect them, I still honestly wonder whether a guidebook is really necessary.

Anyway… I still think minimizing LLM use and coding directly is a fairly old-fashioned development culture and hard to recommend. But I clearly realized that I am not yet at a professional level where I can treat them lightly and tear them down.

Always be humble. Always learn.