Just started vibecoding.

[equinox654]

Had some videos come up about astra last sunday and had work buy me a Premium sub.

Here is my experience from the past 2 weeks:

I had been sleeping on AI. My company is around 3.5 years old and I am a minority partner and the engineering manager, so I am heavily incentivized to be involved with things outside my normal scope. In the past I have made product catalogs, played IT guy and helped out in anything requiring tech knowledge.

We have standard products, but do a ton of custom steel work. We have around 20k drawing pdfs from products we have created and things are getting to the point that we cant find anything. I implemented PDM recently, but manpower wise its impossible to move everything to it, so it will take a long time to be useful.
I had the idea to vibecode an indexer for my drawings. I worked with it for half the day yesterday and ended up with a solution that can find any drawing I want and open its working directory while displaying large thumbnails and contextual information.
That project turned into a whole ass intranet running off of the server next to my desk.
I had it make a page on the intranet where the shop guys request revisions and upload pictures or pdfs and it creates Jira card for my team.

I also had it review our mathcad files and create a desktop application that implements our calculations for quick design checks. - I still need to have my engineer validate it.

The context it gained by having it create the drawing indexer helped me with a solidworks automation test. Our main product is load bearing hatches that are cast into precast blocks for infrastructure.
I gave it an example model and asked it to resize it to another size without using vision to save tokens. It referenced similar sized hatches that exist due to the context from the indexer it created. Initial results are promising. - I essentially make infinite hatch size and hardware combinations all day, so the automation will help.

It also made me a solidworks addon to make generating structural members in the solidworks toolbox less of a pain in the ass.

If I wasn't almost out of tokens I would be having it create a desktop application for automation in as many product families as possible.

It is amazing at checking drawings. We had to make 48 pipe stands for a customer, but we had to change them from welded (in their drawing) to bolted connections. I gave it their drawings and our drawings and explained the context and had it red line our drawings and it found more errors I wouldnt have noticed and pointed out errors in the customer drawing.


Fortunately my team is small consisting of another partner (The Engineer) and my son and a new draftsman. So I can automate the bejesus out of everything and the worst thing that will happen is that we will put the 4090s and 5080s in our workstations to use other than solidworks.

 

[LukeTbk]

I think that one place that it shine the most, doing custom solution for yourself, not try to compete with what a solid team of human programmer can do for general product.

And that a bit why it did not change much the public facing software world to the public eyes, because most use case its peoples/enterprise making custom solution for themselve, that they are never polished enough to sales/share (or not particular want too as a competitive advantage, troubles it opens or so custom-specific there is no one else that would want it)

Know someone that convert house plan into bill of materials for example.

 

[equinox654]

I am so AI pilled now that I had work buy the $200 claude subscription yesterday. I had it code a solidworks mcp server and taught it about my products and standards.
It has been making my drawings for me today. It made a complex hinged trash rack (for sifting water run off) for a infrastructure project that would have taken me a few hours especially since the pdf provided was blurry.
We have made a few sizes of these before and I copied one and opened the main assembly and gave it the pdf and told it that the drawing is for harris county so it might be able to find a clean version online.

It spun for about 25 minutes and spit me out a clean submittal drawing. I couldn't find any errors in its work.
Had it also do some trivial drawings where the customers send over chicken scratch or sales attaches a email thread.

You don't see anyone admitting to cad automation at this scale online, but it works now with solidworks if you vibe code an mcp server.

Edit:
Entry level drafting is cooked.
Claude is handling drafting work better than two of my newer hires.

 

[equinox654]

Started a personal project.
My all time favorite BBS door game: LORD 2. I am giving it a remaster overlay. Got the idea from the Diablo 2 remake.
You can hit tab and turn the remaster graphics off. Its still early, but it is surreal.

 

[Deleted member 314515]

Yeah ... I just landed a 30k job. Building an automated system for my friend's multi-million dollar stucco business. I'm streamlining everything for them so they can sell it for 10 million in a few years. I know I'm supposed to hate AI ... but this is ridiculous. I pay my coding friend to audit / validate everything so I don't do something massively stupid before things ever go live. But I also create 8 trillion back ups of data at all times, just in case.

 

[equinox654]

So I think im getting promoted to be CTO or VP, or some other bs title at my company. I am already a 5% share holder from helping start the company.
All because of the effeciency gains I have pushed out with AI. I recently started doing analytics and automated our mrp data entry and streamlined shipping with a portal that pulls from our google worksheets.

Basically I get to be an executive because I can vibe code. 30 years of being a computer enthusist doesnt hurt either.

 

[LittleBuddy]

Is vibecoding prompting and getting results or more sophisticated?

I do use Grok Build CLI for coding workflows, I use custom skills and custom workflows and have MCP setup for VS, Blender, Photoshop, and a few other tools. I build game mods for games mostly, but have created a video editing app that supports all the features from ffmpeg. I'm also working on a shader injector for Stellar Blade based off the Rebirth Shader Injector. Most the mods I've made are for KCD2 though, haven't released them since I don't feel turning it into an unpaid job. Yesterday I had Grok rewrite a shader for KCD2 without needing to recompile it, it was byte edits, worked perfectly, it used REA with Ghidra to complete the task.

 

[equinox654]

I just think it’s a fun term. I know what I want and how it will work. Like I’m laying out the architecture to Claude and often ask for input. I just don’t know how to code it my self. I do query it to find out how the system works and have it write documentation.

 

[Pianomahnn]

Even though I don't think my definition is correct, I use vibe coding when describing AI coding based on noobs prompting. A person who doesn't write code or understand the fundamentals of writing code.

Vibe coders prompt in a way that produces buggy output.

But that's quickly becoming a non-issue as the LLMs become more sophisticated, with skills and whatever else filling in the Viber's knowledge and experience gaps.

 

[philb2]

A person who doesn't write code or understand the fundamentals of writing code.

I used to code a lot, but in languages like FORTRAN and PL/1. Just try to find a modern Windows PL/1 compiler.



But I don't know even squat about PowerShell. I'm vibe coding PowerShell using Claude AI

Vibe coders prompt in a way that produces buggy output.

Why is that? Lack of prompting precision? Something else?

 

[Pianomahnn]

I used to code a lot, but in languages like FORTRAN and PL/1. Just try to find a modern Windows PL/1 compiler.



But I don't know even squat about PowerShell. I'm vibe coding PowerShell using Claude AI




Why is that? Lack of prompting precision? Something else?

It's a meme. But "make no mistakes" as part of the prompt. It's my hangover bias from my earlier exposure to people's great ideas.

Things are much better now, as you're seeing. Though, I suspect you prompt differently than someone in our sales organization. You know fundamentals.

 

[LittleBuddy]

Why is that? Lack of prompting precision? Something else?

Prompting is just a part of it.

Essentials would be:
- development plan
- specialized skills for the task (e.g., development skill tailored around the API and coding language version)
- an rhai workflow (so the LLM is tailored to how the user wants to work)
- prompt

Optionals that make it better:
- a knowledgebase of the API's that are interacted with
- language/API related skills (e.g., parser + linter)
- advanced skills (e.g., a simulator that simulates the code running on the API)

There are various other ways to improve the outputs.

 

[mashie]

Depending on the model you don't have to be very specific with the prompt.

When I wanted to improve the audio format detection in Kodi my prompt was something like this in Hermes:

Can you investigate if Kodi on the Nvidia Shields can be tweaked to detect immersive audio formats like Atmos and DTS-X? Mediainfo on the main PC can detect it if you need to see a working example.

After that we just had a few back and forwards discussions and Kodi was fixed 3h later and my library had been rescanned to show the correct formats. This was using Deepseek V4.1F locally.

 

[uOpt]

Even though I don't think my definition is correct, I use vibe coding when describing AI coding based on noobs prompting. A person who doesn't write code or understand the fundamentals of writing code.

Vibe coders prompt in a way that produces buggy output.

But that's quickly becoming a non-issue as the LLMs become more sophisticated, with skills and whatever else filling in the Viber's knowledge and experience gaps.

Vibe coding is doing LLM code generation and using that code without reviewing or understanding it. Whether it is done by a n00b or a (former?) professional programmer.

 

[philb2]

Vibe coding is doing LLM code generation and using that code without reviewing or understanding it. Whether it is done by a n00b or a (former?) professional programmer.

That's a great definition. What if LLM code generation is being done by an experienced professional?

A guy I know retired early this year as a senior developer. In the past year or so, his job shifted from reviewing code developed by his people to "junk code" developed by AI.