Prefer to watch? It's all in my YouTube video, it's available in English, Hindi and Turkish: all the projects with the agent from 20:19, and the big project by another developer from 1:18:22

All the materials for the video are on my Telegram channel, come join: t.me/mustangbas

Plan

1. Which projects I built with the agent
2. Steam parser
3. Dribbble parser on requests
4. Steam login
5. Fixing Rabby Wallet
6. Another developer's 6500-action project
7. Conclusion

Which projects I built with the agent

They go from a simple parser to another developer's project with 6500 actions. In short, here's the big picture first: how many actions came out, how many prompts it took and where it stumbled

Steam parser

The first project is a parser for Steam discounts: three categories, filters and a separate folder with a CSV for each category. All of it went into one prompt

During testing the agent ran the project in run mode by itself and saw that the folder had two categories out of three. Then it started debugging and went looking for why one was skipped, and kept debugging until everything worked

It came out at 114 actions, with plenty of extra ones, plus for some reason it creates separate variables for selectors. I didn't touch the script at all. Writing a parser like this on dynamic selectors by hand would take maybe an hour or two

Dribbble parser on requests

Here I needed a Dribbble parser over HTTP that:

• goes into every topic and saves the images
• keeps an image limit
• checks for blocks
• runs in several threads

The first version wasn't fully on requests. The agent opened the site in the browser, grabbed the cookies and only then collected the links over HTTP. By the way, it added proxies for the requests by itself, I never asked for that

With the second prompt I asked it to drop the browser completely. For the test I set 16 images, added a proxy and cleared the folder. The images started downloading, and the browser never opened once

The whole parser took about 30 minutes and two prompts, and came out at 111 actions

Steam login

Here everything is mixed together:

• the browser
• HTTP requests
• the Authenticator module for Steam Guard
• a paid fingerprint

Basically the bot has to log in to the account with a login and password or with a maFile, download a picture from another site and set it as the avatar

In the first version the agent caught a beginner's mistake with the cookie window on its own and fixed it. It never ran the script in run mode, and didn't create a single profile. With the second prompt I asked for proper resources and both kinds of login

It came out at 234 actions. There are extra ones, just not too many. It handled two-factor fine: it gets the Steam time, gets the code from the maFile, and there's very little junk there

In run mode both accounts logged in and set the avatar. The whole thing took two or three prompts

Fixing Rabby Wallet

An existing project installs the Rabby Wallet extension in the browser, creates a wallet and saves the seed phrase to a text file. It was crashing with an error, and I just handed the error to the agent

After the first fix everything worked in record mode. In run mode with two threads both threads crashed right away, and this is where the agent got stuck for a long time. According to the log, this one answer took 2462 seconds, that's about 40 minutes

It fixed one error, a second one popped up, then the first one again, and so on until the end. All I basically had to do was confirm resources. Then I asked it to add logs, the error came back, I sent it again, and after that everything worked

I barely touched anything by hand here. It restarted and tested, by my estimate, at least 30 times

Another developer's 6500-action project

Someone just sent me this project: around 6500 actions and a million resources for a Russian-speaking audience. Just figuring out something like this by hand takes at least 5 hours, which is why on freelance sites people turn down projects like this in 90% of cases

First I asked it in Thorough mode to briefly explain what this project even is. It broke it down pretty well overall:

• a farm for crypto airdrops
• 161 functions and about 6400 actions
• wallet registration, working through projects, and separately Telegram projects

Then I asked it to sort the resources into tabs and make Russian and English localization. In about 20 minutes it sorted everything out, translated every resource and added a description to each. By hand that's 30 or 40 minutes of heavy routine, so I rate this part at a full 100%

With descriptions it's better to state in the prompt right away which resources need them. I didn't, so it added them everywhere

Next I asked for a plan to rework the whole project. The plan came out really huge, by hand that's three days of monotonous work. I took three tasks from it:

1. Rename functions and variables
2. Bring functions to one format and add a description to each
3. Add more logs with the Logger module and explain the heavy loops

After 4 hours the agent sent a report: the first task at 100%, the second at about 85%, the third at 70%. I checked the renaming and the descriptions, everything is in place, and there are no old names left in the project

The logs turned out more interesting. The agent itself wrote that it covered the big functions only partially. I went into the functions where, by its own words, it had added logs, and didn't see new logs there either. Most likely it lost context and counted the logs a person had written before it as its own

Plus it often stopped in the middle of the work and had to be restarted. Why that happens I'll cover in the article about context. Across the whole project I'd rate the result at around 40%, maybe 50

Conclusion

Honestly, at first I thought the agent was still raw and would mess up in a lot of places. By the second project I already understood I was very wrong

What stuck with me most is that it doesn't give up. It hit an error, went to fix it, hit the next one, fixes again. It built the Steam parser from one prompt and fixed Rabby Wallet completely. It does pile up extra actions and variables, and it sorts resources its own way in places

I'm sure it would have pulled off the big project too, with everything split into stages with a plan and a decent prompt

The downsides the developers still have to work on:

• there's no auto-confirm, everything has to be confirmed by hand
• on a big project it loses a lot of context
• tasks take a long time