Adventures coding with AI

[Drizzt321]

It never fails, just when I think things are amazing... AI is the dumbest smart fresh out of masters CS first job idiot. Occasional flashes of brilliant tying things together, and then just cannot seem to wrap it's weights around and slightly expand/contextualize the pattern I'm trying to get it to use.

 

[hanser]

Opus 4.5+ writes far more code comments than GPT-5.5. Most of GPT's comments are doc comments, whereas a huge percentage of Opus' are inline. Broadly, these CLI's (except salesforce) were developed using the same workflows.

 

[w00key]

It never fails, just when I think things are amazing... AI is the dumbest smart fresh out of masters CS first job idiot. Occasional flashes of brilliant tying things together, and then just cannot seem to wrap it's weights around and slightly expand/contextualize the pattern I'm trying to get it to use.

Haha. Although I must say that in my financial reporting and restatements today (adjusted P&L, to see wtf actually happens, not what is tax optimal with intercompany flows) Opus 4.8 just gets it?

It is also good at flagging issues like hey, energy increased a lot, you sure, to which I answered, yeah our fixed price period ran out and this little thing in Ukraine fucked the market. Or missing intercompany charges, accountant messed it up, needs moving to right GL account.

It's way less boring and far more productive than digging through the numbers alone.


But yeah, pretty good most of the time then you see a pretty dumb thought in the trace, then it goes no wait, and corrects itself. Effort = medium works well for this job.

 

[Zich]

Opus 4.5+ writes far more code comments than GPT-5.5. Most of GPT's comments are doc comments, whereas a huge percentage of Opus' are inline. Broadly, these CLI's (except salesforce) were developed using the same workflows.

This is good to know. Opus's inline comments really get on my nerves. They are almost always gigantic, even with urging to the contrary.

 

[w00key]

Found on Reddit:


View: https://www.reddit.com/r/ClaudeAI/comments/1tucsfe/i_had_opus_48_build_temu_league_of_legends_in/

Fun weekend project just to test out 4.8 against a pretty vanilla setup.

Started out with a simple prompt, "build a temu league of legends, web-only with online, room-based multiplayer".

It built it... fully functional in one shot! Then I went kinda ham and started iterating on every aspect little by little. I made it spin up subagents to each focus on designing a character. Spun up subagents to focus on designing abilities and their SFX/VFX. Did a few passes over the map, mobs, minions, etc.

Then I put Ultracode Workflows to task and, across a handful of prompts, had it optimize performance, balance, and a variety of other things. Workflows is insane btw...

I also used /goal pretty heavily as well. I'd sit and play and make note of like 10-15 tweaks, bug fixes, etc at a time and just let it go to town.

4.8 is a one shot machine... 5.5 ain't doin this...

Anyways, figured I'd share Happy to answer any questions in the comments.

Try it out solo (with bots) or with friends! https://lmaomoba.com

Whoa damn. This is pretty insane... OP also wrote

The very first prompt was, exactly:

"/goal I want you to build the temu version of League of Legends in this project. I dunno what LMAO stands for, but we need to come up with something. It needs to be built with next.js & react and has to support online play with others. It should support joining a host's room like a lot of modern multiplayer web games. You have full autonomy to build this however you see fit aside from the next.js & react constraint (that is purely for deployment though... use whatever tech stack can fit in that deployment model; I just don't want to think about or deal with that side). The game has to be fully featured. You may execute this entire project without any intervention from me. Don't ever stop for my input until you're done and ready for me to review the full site, portal, game, etc. Have fun! Be creative! You're an expert game designer and developer, so use all the knowledge and tactics at your disposal to make this thing a hit."

That alone literally got me a working, playable POC. I think it was something like 4-5 of those /goal prompts I mentioned before it actually felt pretty good. Then it was mostly expanding the content, polishing, improving controls, and other refinements.

 

[w00key]

Hahaha and a spinoff


View: https://www.reddit.com/r/ClaudeAI/comments/1tuopn2/my_attempt_at_a_1_prompt_temud_game_temu_version/


On the $20 plan. What the hell lol.

goal: I want you to build the temu version of Anno 1602 (1998 Max Design game) in this project, but with a more modern graphic choice. I dunno what fun name we could use for the project, but we need to come up with something. It needs to be built with next.js & react and has to support online play with others. It should support joining a host's room like a lot of modern multiplayer web games. You have full autonomy to build this however you see fit aside from the next.js & react constraint (that is purely for deployment though... use whatever tech stack can fit in that deployment model; I just don't want to think about or deal with that side). The game has to be fully featured. You may execute this entire project without any intervention from me. Don't ever stop for my input until you're done and ready for me to review the full site, portal, game, etc. Have fun! Be creative! You're an expert game designer and developer, so use all the knowledge and tactics at your disposal to make this thing a hit. Would be cool if public reaction would be "Holy shit! A.I. made THAT?!".

 

[hanser]

Microsoft says hi

https://microsoft.ai/models/mai-thinking-1/

 

[Drizzt321]

Huh, interesting. Although I don't trust it, it's built by MS, when was they last time they actually built something worthwhile, than just more crap on top of their monopolistic lock-in?

 

[w00key]

Microsoft says hi

https://microsoft.ai/models/mai-thinking-1/

Ooh, feels very Claud-y with a focus on safety. Gemini is overly yolo and OpenAI has super weird refusals, so it's a great thing they do this. One thing I like with Claude is that it is very "aware", prompt injections rarely work and it's almost like it is ethical? Not sure what word to use to describe it, but we need more of this.

 

[hanser]

Huh, interesting. Although I don't trust it, it's built by MS, when was they last time they actually built something worthwhile, than just more crap on top of their monopolistic lock-in?

Uh… what monopoly does Microsoft have… in literally anything?

 

[Drizzt321]

Uh… what monopoly does Microsoft have… in literally anything?

They're certainly monopolistic in behavior, especially when it comes to various "Enterprise" and office software stuff. Speaking of which...weren't they actually convicted of being a monopoly at one point in time? Has the leopard actually changed it's spots?