When you can kind or speak, you may in all probability vibe code. It is actually that straightforward. You merely talk your concept to the AI chatbot of your alternative with pure language, and it’ll get to work. Whereas all tasks could not find yourself being a pearly gem and even totally purposeful, the hassle concerned to create an online app by simply chatting is minimal.
I’ve performed round with vibe coding fairly a bit, creating random tasks right here and there. I’ve principally tried it out for proof of idea or simply chatbot testing functions, however hardly ever in an try and make one thing essential or purposeful to the purpose the place I might see myself utilizing it each day. Then I started in search of an e-reader machine and walked away pissed off by how extremely “dumb” most of them are, particularly Kindles. Extra on that later.
Whatever the motivation (however principally in defiance of Amazon), it prompted me to attempt to vibe code a repair, however with a twist. I wished to see if I might vibe code a purposeful e-reading utility with all the options I wished. Figuring out that even when I obtained this factor up and working, I probably would not use this each day, however I wished so as to add some aptitude.
The query grew to become what AI chatbot to make use of. I did not hassle attempting to decide on. I used Gemini, Claude and ChatGPT to create what I wished after which see if one produced higher outcomes.
The immediate
Evaluating chatbots is difficult to do, particularly when attempting to imitate the identical dialog or vibe coding mission. Consider me, I’ve tried. I wished to ensure all the instruments I examined have been utilizing the identical immediate, however first, I wished to refine that immediate to get the most effective outcomes, so I got here up with a method to get me there.
First, I constructed your complete mission at its most elementary stage with Gemini. As soon as I favored the place the mission was (profitable and purposeful proof of idea), I requested it to create a immediate in order that I might add it to every other chatbot. Gemini generated the immediate, I saved it as a file, and I uploaded it to Claude. I went via this course of once more, permitting Claude to catch and sort things I hadn’t thought of when constructing the mission and Gemini. As soon as that course of was full, I requested it to create one other immediate so I might add it to ChatGPT.
The thought was to have all three chatbots have some enter on the precise creation of the mission and, in flip, the ultimate immediate. As soon as the immediate was created, I uploaded it to all three chatbots in a separate chat to see how persistently they carried out.
The mission: The Tome Reader
This mission was born out of my frustration with Amazon’s Kindle gadgets. Anybody who likes to learn and take heed to their books can achieve this with real-time highlighting within the app for iOS or Android, however in spite of everything this time — practically 20 years — you may’t do that on a Kindle. In reality, it wasn’t way back that customers gained the power to realize real-time textual content highlighting with the assistive reader enjoying, which is so near the app’s performance. As of proper now, you may solely learn or take heed to an audiobook on Kindle, not each, which is laughable, and so was the thought of Amazon proudly owning all my books. I obtained to pondering that I might simply vibe code an answer. I name it the Tome Reader.
I wished to create an immersive e-reader net utility that might learn your books aloud to you (with real-time textual content highlighting, after all) whether or not the textual content was pasted or your personal PDF or EPUB file was uploaded. Along with studying the textual content aloud, the net app would create background music relying on the content material of the textual content in a subset of classes (impartial, gothic horror, sci-fi, nature, fantasy, underwater, western, thriller), and generate further sound and visible results when sure set off phrases have been spoken in real-time. The complete mission was created in a single HTML file so it might run inside an online browser with out further dependencies.
Constructing with the chatbots: The primary spherical
Gemini
Gemini made all of the options I wished from the Tomb Reader with relative ease.
Gemini allowed me to determine how far I might stretch the perform of this net app, and thus a majority of its performance comes from Google’s chatbot. It allowed me to hash out some small points to start with, which prevented the TTS voices from loading. As an alternative, it created an initialization display screen that might pressure the voices to load after clicking on an “open” display screen to the appliance. With out any such know-how, the mission would not have gotten off the bottom.
Slowly however absolutely, the performance of the mission started to develop. As a result of stay sound results for sure phrases could possibly be distracting, I made certain so as to add the choice to show these sound results and the background music off. After I obtained a base of the appliance working, I requested Gemini to create a immediate I might share with different chatbots so I might construct it elsewhere if I wished, and that is what it did.
Claude
Claude’s mission gave me probably the most success in some areas and extra hassle in others, however is my private favourite of the three take a look at tasks.
Claude made implausible refinements to the underlying perform of the set off phrases on this mission. Claude prolonged the vocabulary and elevated the visualization when a set off phrase was spoken aloud. That mentioned, Claude made a name that I did not ask it to, although the logic certainly made sense.
Initially, I believed the mission wasn’t working as a result of after I went to check its performance, simply the primary set off phrase would create the specified impact in a string of practically 10. It took a while for Claude to lastly reveal it had determined to solely enable the sound and visible results set off as soon as per sentence in order to not “spam” the person. This made a variety of sense, however the mission was extra proof-of-concept than purposeful reader, and Gemini and ChatGPT generated sound results for each key phrase, which was the anticipated performance.
All that mentioned, there was no particular instruction throughout the immediate for the variety of instances the sound and visible results performed. Whereas it wasn’t essentially what I wished, I did admire the consideration of the general person expertise to make such a name. Then in spite of everything of these refinements have been coded in, Claude then up to date the immediate, and I took it with me to ChatGPT.
ChatGPT
ChatGPT did not create further options after I requested it to at instances, however nonetheless managed to recreate the mission completely after I gave it the ultimate immediate.
By the point I had created the up to date immediate with Claude, there wasn’t a lot else I might suppose to do after I uploaded it to ChatGPT. Fortunately, OpenAI’s chatbot created the mission with ease, regardless of being the slowest at producing the code. The one perform I did ask ChatGPT so as to add to the mission — to create a devoted quantity slider for the background music so it could possibly be turned off fully if one simply wished a devoted e-reading expertise — failed persistently. Ultimately, I went again to Claude to ask for this performance and recreate the immediate.
Spherical 2: Recreating the identical mission
Regardless of utilizing Gemini 3 Professional to initially construct the mission and the free variations of ChatGPT and Claude, all three created the mission, however not with out points. I had spent most of my time inside Claude refining the mission, and it was accountable for creating the ultimate model. So it was extremely stunning to seek out out that when importing that immediate into a brand new chat, the mission would not load previous the primary “initialization” web page. Regardless of having no points in any respect with earlier iterations, it took 11 (sure, actually) further full rebuilds to determine what was happening.
Recreating the mission with each Gemini and ChatGPT labored flawlessly. All features, primary and superior, labored as they need to, together with file importing, take a look at highlighting, text-to-speech output, and each audio and visible results when set off phrases have been spoken aloud. Going again to the fashions, I noticed little or no distinction in perform or efficiency when giving the identical immediate to every of the chatbots.
Chatbot inconsistencies
Buying the file to check was all the time straightforward with Claude. Not solely did it provide a preview of the mission so that you by no means wanted to obtain the HTML file in any respect, however when you wished to (which I usually did for testing), it was made obtainable to obtain instantly. This selection was solely typically supplied with ChatGPT, whereas at different instances I might solely copy the HTML and reserve it by myself.
Regardless of having the fewest errors and total qualms with Gemini, it all the time required you to take the lengthy route and do it this fashion. All that apart, the truth that ChatGPT would solely typically provide to permit me to obtain the file as HTML instantly was peculiar and somewhat irritating.
The winner: It relies upon
Defining a winner for any such take a look at is hard, as all chatbots have execs and cons to them. In a way, all of them win. Every was in a position to create a purposeful model of the mission sooner or later, nevertheless it usually took repeated efforts.
In the end, the winner is the person. It goes to indicate that, whatever the mannequin getting used, a stable set of directions can get you far. I used to be unable to tell apart variations in efficiency or perform between the app created by Gemini 3 Professional versus the free variations of ChatGPT or Claude.
This really goes instantly in opposition to what I discovered when having an identical dialog with each the Gemini professional and free fashions. Whereas that was one other day, one other mission, and one other mannequin, it goes to indicate {that a} stable immediate can get you extremely far on the earth of vibe coding.
