What I built with Replit in 1 month
Many things are half-baked :P
First on Catalan at: En Rub aprenent.
A website to keep track of my digital collection of Pokémon cards. I already told you about this in the newsletter and went a lot further, I'll summarize the news at the end.
A Habit Tracker to quit smoking (not for me :p). I haven't finished it yet, it's given me more headaches than I thought for being the kind of "absurd app" that everyone is building. I think my initial prompt was too demanding, and I had to correct it quite a bit...
A website for tracking and trading crypto with Hedera. Don't put red laser eyes on me yet. It was a proof of concept to encourage a more "crypto-bro" friend to use Replit and see if the chili could get into this garden. We discovered that the knowledge of the model towards the libraries of this was very old, and it didn't work out. It quickly monitored values on a testnet, but trading… no more. I don't have it published, either.
A web-crawler to extract documentation on web accessibility WCAG from the W3C. A bit regular, to make crawlers you have to give it good examples of the HTML that you will “parse” and I wanted to trust it to see if it was more capable than using LLM directly. Spoiler: it is not. Although LLM dresses in silk, LLM stays. I had to use ChatGPT to fix “the mess” and everything went like, ahem, like silk, until the W3C blew me away for a supposed DDoS attack :p “they just don't have patience”!
A portal to consume RSS and create summaries with AI. I didn't give it a try, I was too lazy.
Updating the Telegram bot of a channel I participate in a lot, Mossegalapoma , was a total disaster. I'll leave some comments about that later as well.
MediaWikiConverter, the tool to extract data formatted like MediaWiki to JSON as a support tool for Pokémon card tracker. As an internal tool, I don't have it published.
A “visually beautiful” product website that talks about the goodness of Replit. It was a quick test to see Replit's knowledge of itself and how it put together a typical product website. Not bad, not great. I didn't publish it, either.
A form of 7 questions, 7 pages 1 per question with, 6? Answers where you have to spend 10 points between the answers. It is the Belbin Roles test and ends with the results, a radar diagram and the possibility of exporting the results to .csv or .pdf. I messed up doing the deployment, there were some parameters that were missing, I don't know if it was a Replit error, or I deleted them without realizing it and from that Replit screwed up again and again, until by turning it around with o1, we found what was broken.
Well, it's no small feat, especially considering that I've been on vacation for 70% of the time, sick xD.
I ate the $25 and about 4 more dollars of extra credit, which I think they forgave me.
More info about the projects…
Pokémon cards
I switched from using Postgres to SQLite, then when I saw that the DB was obviously dying when the container was turned off (that's what the cloud has, congratulations son), I ended up with SQLite for development and Postgres, again, for production.
Replit has not solved, nor even proposed, as far as I know, the management of environments, except that you can play around a bit with Git (which you can).
I put unit tests on it, serious mistake, It "touched" the Workflows too much (which would be like different starting points of the application) and then the tests always initialized SQLite when "compiling" the project. Laziness to fix it and not wanting to spend more credits, made me touch the Workflows manually, and now I have a real project, that is, with expired tests that are not used hahahaha!
I asked him for a Readme.md, very well.
Then I made some big, non-trivial refactors, including the BD schema, because I had to add a new card collection and make the cards multi-collection, another congrats to me, I could have done it from the beginning, but it went quite smoothly.
Telegram bot
The issue is this, I have legacy code for the bot. I thought: “well, I'll give it the code and tell it to update the api/framework version, and I'll be 80% done”. Error!!
I don't know if this is or will be the norm, but I found that to use the Agent you had to first set up a project with the Agent, put your code in it and then the Agent would take responsibility... nothing at all.
Okay, we'll spend $0.25 and I'll put the code in later. Cool, plus "making a Telegram bot" is one of the use cases promoted by Replit. Result? It wasn't able to screw up a basic Telegram bot until a few iterations... really?
Then I uploaded the original code and told him to look at it and copy the new API functionalities. I think he looked at me with a blessed face saying “are you talking to me?” and snort… I want to take it back, let's see if we can get it up and running, but total disappointment. Maybe the task is not trivial, but if you see how it made a mess just to set up a “hello world” bot… the AGI one, that's it…
Summarizing
I'm in a love/hate relationship with Replit right now. Still, the potential is incredible, I have a lot of faith in it, especially as reasoning models get better, cheaper, and faster. If we lower the cost of inference, platforms like Replit will be able to give the model more leverage to generate better answers and make it work for them.