I’m one of those people whose knee is constantly jiggling. Especially when I am sat in ‘receive’ mode in a meeting or something. To reduce the jiggling I fiddle with things and the thing I have been fiddling with will be familiar to anyone who likes to see what all the fuss is about with new tech. I’m always asking myself- novelty or utility? (I had my fingers burnt with interactive whiteboards and have been cautious ever since). You may be interested in the output of Perplexity’s ‘Comet’- the browser based AI agent, the outcomes of which are littering LinkedIn right now- or the video below which is a conversation between me and one of my AI avatars… if not either of these I’d stop reading now tbh.

In the image below is a link to what I instructed using a simple prompt: “display this video in a window with some explanatory text about what it is and then have a self-marking multi choice quiz below it.” [youtube link]

It is a small web application that displays a YouTube video, provides some explanatory text, and then offers a self-marking multiple choice quiz beneath it.

Click on the image to see the artefact and try the quiz

The process was straightforward but illuminating. The agent prepared an interactive webpage with three generated files (index.html, style.css, and app.js) and then assembled them into a functioning app. It automatically embedded the YouTube video correctly (but needed an additional prompt when it did not initially display), added explanatory text about the focus of the video (AI in education at King’s College London), and then generated an eight-question multiple choice quiz based on the transcript.

The quiz has self-marking functionality, with immediate feedback, score tracking and final results. The design is clean and the layout works in my view. The questions cover key points from the transcript: principles, the presenter’s role, policy considerations and recommendations for upskilling. The potential applications are pretty obvious I think. Next step would be to look at likely accessibility issues (a quick check highlights a number of heading and formatting issues), finding a better solution for hosting and then the extent to which fine tuning the questions for level is do-able with ease. But given I only needed to tweak one for this example, even that basic functionality suggests this will be of use.

The real novelty here is the browser but also the execution. I have tried a few side-by-side experiments with Claude and in each the fine tuning needed for a satisfactory output was less here. The one failed experiment so far is converting all my saved links to a searchable / filterable dashboard. The dashboard looks good but I think there were too many links and it kept failing to make all the links active. Where tools like notebook LM are offerring a counter UX to text in; reams out LLMs of the ChatGPT variety, this offers a closer-to-seamless agent experience and it is both ease of use and actualy utility that will drive use I think.