Yesterday, we were discussing when we could actually let AI agents auto-merge and publish code.
My take - probably not so soon.
My take - probably not so soon.
These screenshots are exactly why. It is the output from Claude Code when I asked it to add a feature.
Before you think it is a prompt and workflow issue, let me tell you that I have a comprehensive workflow that spawns 9 subagents to review code based on 9 best practices. Two of the subagents review React Native and UX/UI.
Yet, it still keeps making these mistakes:
-
UI does not respect screen safe area - the navigation bar overlaps with the status bar.
-
Layout misalignment issue.
LLM output right now is like a slot machine. Most spins land fine. But about 10% to 20%, it fails at the simplest layout.
The code compiles. The app runs. But visually, it is off.
For web apps, there is a test harness. You can now let Claude Code see your browser UI through extensions. It spots visual issues and self-corrects. This helps a lot.
But for mobile apps? I still haven’t come across a way to let Claude Code see the screen.
And that I think is the real bottleneck in AI coding now. I spent more than 50% of my time doing manual UI testing.
AI can write code that compiles. But right now, it can’t tell if the screen looks right.
Until an AI agent can see its own work the way a developer does, we probably still need a human UI tester.
Enjoyed this? Subscribe for more.
Practical insights on AI, growth, and independent learning. No spam.
More in AI Agents
Your AI agent will get prompt injected sooner or later, because it is easier than most people thought.
Most people think prompt injection needs a carefully crafted adversarial prompt by an experienced hacker. It does not. Someone who understands how LLMs work ...
Can LLMs like ChatGPT do reasoning? It failed my casual tests in under 30 minutes.
… more
What’s the most common hallucination you've seen from an LLM?
For me, it’s when you ask, “How do I do X in Y app or software?”
Recently, I heard an interesting view from Jeremy Tan during a panel discussion.
It sounded absurd at first, but on second thought, I think it might actually happen.
I’m honored to be invited to moderate an insightful roundtable on 𝗘𝗹𝗲𝘃𝗮𝘁𝗶𝗻𝗴 𝗙𝗼𝘂𝗻𝗱𝗮𝘁𝗶𝗼𝗻𝗮𝗹 𝗦𝘆𝘀𝘁𝗲𝗺𝘀 𝗶𝗻 𝘁𝗵𝗲 𝗔𝗜 𝗔𝗴𝗲, hosted by The Ortus Club and MuleSoft — with an exceptional group of tech and data leaders across industries like banking, telco, healthcare, transport, finance, and travel.
We unpacked tough questions on:
Last Friday, I ran the Foundations of Claude workshop for Sequoia Group, a leadership and OD...
As Sequoians, they believe in going high tech so that they can go high touch. They have been using AI, but they would like to learn how to use it better.
Your AI agent will get prompt injected sooner or later, because it is easier than most people thought.
Most people think prompt injection needs a carefully crafted adversarial prompt by an experienced hacker. It does not. Someone who understands how LLMs work ...
Recently, I heard an interesting view from Jeremy Tan during a panel discussion.
It sounded absurd at first, but on second thought, I think it might actually happen.
Last Friday, I ran the Foundations of Claude workshop for Sequoia Group, a leadership and OD...
As Sequoians, they believe in going high tech so that they can go high touch. They have been using AI, but they would like to learn how to use it better.
Can LLMs like ChatGPT do reasoning? It failed my casual tests in under 30 minutes.
… more
What’s the most common hallucination you've seen from an LLM?
For me, it’s when you ask, “How do I do X in Y app or software?”
I’m honored to be invited to moderate an insightful roundtable on 𝗘𝗹𝗲𝘃𝗮𝘁𝗶𝗻𝗴 𝗙𝗼𝘂𝗻𝗱𝗮𝘁𝗶𝗼𝗻𝗮𝗹 𝗦𝘆𝘀𝘁𝗲𝗺𝘀 𝗶𝗻 𝘁𝗵𝗲 𝗔𝗜 𝗔𝗴𝗲, hosted by The Ortus Club and MuleSoft — with an exceptional group of tech and data leaders across industries like banking, telco, healthcare, transport, finance, and travel.
We unpacked tough questions on: