wrote: January 22, 2026
You might have heard people say "AI is going to take your job", especially for developers. Is this true? Will AI be able to fully function as a human engineer and take our jobs? One way to find out is to put that to the test — and that's exactly what I did.
I've been fascinated with tools (one of the reasons I studied computer engineering, which turned out not to really be about that, lol), but AI tools are something even more interesting and unique — they aim to automate and replace what humans can do, even with complex tasks, in an unprecedented way.
Table of Contents
- Introduction — what tools, what tasks, my objectives
- Method — what I did to accomplish that
- Results — what AI agents gave me
- Conclusion — my takes on this and limitations
I. Introduction
The goal of this experiment is to see how AI agents perform in terms of UI generation and code implementation for web frontend tasks. As a frontend developer with no design background, I can't work without guidelines from the UI/UX team. While waiting for the design team to hand over the designs, I tried to interpret the flow using AI and asked it to generate UI mockups — that way, I could continue the work.
What turned out wrong is that my OCD brain couldn't stop throwing task after task to various AI agents, one after another, and that continued for a while.
For context, this is not my first attempt at this type of comprehensive test on AI agents. The first one judged code implementations, which I found quite interesting but failed to document. So this time I'm determined to write it up. Let's dig in.
Tools and Tasks
AI Agents in Chat Form and on Terminal via CLI
(Part II: AI on Terminal — might or might not write about it)
| Tool | Provider |
|---|---|
| Claude + Claude Code | Anthropic |
| ChatGPT + Codex | OpenAI |
| Gemini Pro + Gemini CLI |
UI Mockup Tools
- Stitch
- Manus
Full App Mockup (Prototypes)
- v0
- Base44
- Replit
- Grok (could fit in the chat category, but it's capably limited in my opinion — that's why it's here)
Note: There are several other well-known tools that are probably highly applicable for this test, such as Lovable and Gorilla.
II. Method
- Device: 16-inch MacBook Pro M1 base model
- Internet Speed: 68 Mbps – 600 Mbps
- Prompt: All AI agents received the same prompts
Step 1: Interpreting the Flow
I asked Claude (in chat mode) to interpret a flowchart originally in Korean into English and describe it in words. Then I asked it to generate prompts based on what was produced. I have no prompt engineering skills — it was based on personal experience, so the commands were pretty straightforward. I also wanted to keep things as simple as possible to reduce complexity in result analysis.
Step 2: Generate UI Mockups
The initial goal was simply to generate UI mockups I could use to develop the project. I started with the recommended tools — Stitch and Manus — which I wasn't familiar with. Then I moved on to the ones I'm used to: Claude, Gemini, ChatGPT, and Grok.
Step 3: Generating Full Apps
For full app generation, I turned to v0, Base44, and Replit. I also used Grok, but was quite disappointed with it.
III. Result
The metric I used to judge performance was the end result only — I did not factor in things like generation time or additional features each agent offers. Results are subjectively examined based on my experience and personal preferences, so biases may appear. I'd love to hear your opinion too.
! Disclaimer: While screenshots would help illustrate the outcomes, I cannot include images of the results due to privacy reasons.
UI Mockup Tools
Worst to best:
** Manus** — The weakest performer, quality-wise. The design was obviously AI-generated, with weird characters, incorrect spellings, and an off color theme. Instantly, I knew this wasn't something I could use. The one upside: it's quick.
** Stitch** — A bit better, but still very basic and inconsistent. It generated several page mockups from one prompt, but they looked incoherent. It might be acceptable for simple use cases like a single-page website or a basic app.
Full App Generators
Best to worst:
** Base44** — Quite impressive. With the exact same prompt, it produced a nice interactive app on the first try — gave me the "it understood the assignment" vibe. Another strong point: the code structure and implementation were nicely separated.
** Replit & v0** — Too close to call. Both are quite capable and offer solid additional features. Replit lets you choose between app or web; v0 has a design mode among others. The generated apps from both look good and function well.
** Grok** — Out of the competition. Not because of any bias, but the results simply weren't up to par.
AI Chat + CLI Tools
Last but definitely not least — the giants:
Claude Code is the best in terms of coding. ChatGPT ranks second. Gemini comes last.
At one point, Claude Code showed some weird garbled characters in the terminal — hopefully I didn't break it.
IV. Conclusion
My Takes
- Some AI agents are really good; some are pretty trashy.
- At this point, I don't think AI is capable of taking developers' jobs — at least not yet — especially when human elements are involved (the kind of nuanced judgment that is computationally tedious and expensive for AI to replicate).
Limitations
This experiment required constantly prompting various AI agents, switching between many tabs and windows to check on each one. In that condition, I'll admit the process was probably not as thoroughly examined or fairly judged as it could have been. Despite that, I put in significant effort and time to be as fair and sensible as possible.
Written January 22, 2026 — finished while also working