AI ML
Post #7392, on Nov 7, 2025 in TG
Four Hours Past Bedtime Still Debugging AI Prompts
Description
The meme uses a distorted, smirking Mona Lisa face with the top text reading 'WHEN IT'S ALREADY 240 MINUTES PAST BEDTIME AND YOU'RE PROMPTING "STILL BROKEN"'. The Mona Lisa's expression has been edited to look exhausted yet smugly determined, with dark circles and a knowing smirk. It captures the relatable experience of staying up way too late trying to get an AI/LLM prompt to work correctly, iterating endlessly on something that remains broken. The imgflip.com watermark is visible in the bottom left
Related Memes
Use J and K for navigation
Comments
23Comment deleted
240 minutes? That's only 4 prompt iterations. The real pain starts at iteration 47 when you realize the system prompt you've been editing was cached the whole time
That's the face you make when the AI has a state management issue, and you've been reduced to a state management issue
You’re absolutely right Comment deleted
Am a supposed to empathise with vibecoders, lol? Comment deleted
And this point you're not developing anything, rather just gambling Comment deleted
its better to write 100 lines of code yourself then let ai generate 1000 lines of unmaintainable boilerplate with a lot of crutchs Comment deleted
Delegate your work. Sure, AI often is just a junior dev who still needs guidance, but it's way faster than a junior 😅 I've long passed 20 years of coding. And yet I let AI generate most of the code for me for years now (at least as long as it doesn't get too complex). Reviewing what was generated is mandatory, and yes, it's not perfect, but man does it safe keystrokes 😄 It's a tool and you gotta use it as such. I'm curios, did you guys stop using AI for coding? Or why does AI hate seem to increase lately? I get the impression, that the general opinion drifts to "using AI equals vibe coding". But isn't using AI only vibe coding when you drop coding standards and / or have no clue what that generated code actually does? Comment deleted
I usually think for hours before i give the actual prompt, in which case it usually does what i need it to do, because i produce a point-by-point spec of what should be done With old models this didn't save any time due to garbage output, new claude models do this well enough. Then I take this boilerplate and extend from there manually Comment deleted
Ok, I see now. You are talking about boilerplate code for a whole new project. I'm talking more about in-IDE coding assistants like Copilot that assist you with existing projects in small blocks of code. Comment deleted
I turned off that glorpshit, distracts me like fucking hell and interrupts my flow Comment deleted
For complex coding tasks I get that. Algorithms often get messy. And for that I do the same, but for that simple stuff too? Comment deleted
Simple stuff gets handled by the ide itself via symbol search Comment deleted
I'm talking beyond getters and setters. Comment deleted
if it's beyond, then it'll disrupt my thought process with shit i didn't ask for if it was reading my mind for prompts, sure Comment deleted
Yeah I feel you 😄 To me that's fine. I start with a meaningful name and comment for my class / method / function / block. And then I see what comes up, maybe look at other solutions / suggested code snippets and fix / adjust them to my liking. Take cases for example. Those may or may not have simple case statements that IDEs just can't generate without frustrating me (for the clear lack of intelligence). A junior can do this task. And so can AI. = Less code I have to write. More time for what actually challenges my brain. And if a comment or the name of a function does not produce usable code for a seemingly easy task, then I'd argue that this kind of prompt (= the comment / function name itself) might need fixing in the first place as it seems ambiguous. I know that it's not always true, especially for less know frameworks (svelte 5 anyone?), but most of the time it is. Prompting is art. Or science. Or both. Comment deleted
My main problem stems from the fact that when i'm writing code, I'm operating in output mode. When I have to read and verify the suggestions, it switches me to input mode. This is what disrupts me the most. Comment deleted
Ok, yeah, I can understand that. You mentioned symbol search. Did you mean actively searching for the symbol or do you start typing the symbols name and let the IDE autocomplete the name? And if you are using auto complete, is that less distracting because it's less code or are there other reasons? I'm just curious and try to understand my fellow devs issues. Maybe there is a problem to solve. Guess that's occupational deases 😅 Comment deleted
usually when i start typing - intending for something specific to show up that i know i'm gonna choose to save keystrokes (for autoimporting too) Comment deleted
Ok, gotcha. Thanks for the explanation 🙏 Comment deleted
also svelte 5 is cool and based Comment deleted
That is what I use the AI auto completion from IntelliJ etc for Comment deleted
That's because your agent is running locally in cursor now Comment deleted
i like how this is also ai-generated Comment deleted