About Disposable AI Interfaces

New year, but posts are still about AI. Another cool use case for agents has been discovered.

Several times recently I've found myself in a situation where I realized that giving the neural network the final decision on some tasks isn't worth it and I need to double-check its work. I understand that it doesn't have enough context for this, but I do. And it's not just one thing to check, but a whole list of different items.

Let's get closer to specific examples.

• Suppose I have some random icon in my project and I want to pick the closest analog from the design system.
• Or I want to assign some coefficients to a data array — for example, task priorities, to make them more convenient to work with later.
• Or I'm preparing data for some pet project, asked the agent to populate the database with information and images from the internet, and I want to validate this and discard the bad stuff.
• Or maybe I'm choosing icons in my pet project and I'd like to look at the suggested options instead of it inserting the first one it finds.

And all of this sometimes needs to be done in a loop many times. For example, for a bunch of rows in a database.

For these cases I want something like a questionnaire. To make a decision you often need to open things side by side, compare. One image, another, you can't see the difference right away, you need to somehow overlay them. You'll get confused doing this. Or, for the other example, typing numbers into a simple Excel spreadsheet isn't very fun, and you can't keep everything in your head. You'll make mistakes too.

And so I thought, why not make this questionnaire dynamically for the task. A micro-application where you go through each item separately and make decisions — choose answers from those suggested by the agent. Options don't have to be limited to two, somewhere there might even be multiple choice. It can even highlight what it thinks is correct so you can work from that. It can help you render all this to make it more convenient to look at. It shows you progress. As a result, it can itself generate a report with answers, which it then processes.

Essentially it's the same wizard as the one built into Claude Code with answer options, but on steroids. With any customization that's needed. 🕶

This symbiosis, where it can create one-time automation out of nothing and I can think — that's the best. I would never write such automation in my life, it would never solve the task properly without me. But once you give it a special interface for the task to help you — it immediately becomes easier for you and it gets clear feedback from you.

I really love abusing HTML for any visualization with the help of agents. From some presentation for the team, to a utility for solving my specific little task.

Everything described as a prompt here. I even made myself a custom command to play with this tool to my heart's content. I attached a video example of what it generates in the comments.