macOS Vibe Coding + Private Science work? Can Vibe coding help with fast tool building?

[Sosab123]

I first got into programming when SWIFT was launched, I wanted to build tools for research because that will inevitably be required. SWIFT seemed a lot more accessible to me at the time as a programming language, but, due to my constraints, I wasn't able to get very far. I did however meet a lot of great people and really enjoyed my time in the development sphere and got set up for a very nice time moving forward.

I am back now, and I have the necessary investments that can allow me to spend the time building the tools that I needed to start the research i want to do without risking my way of life.

I don't plan to sell or add apps to the market, I am just looking to build the tools I need for my research.

I was going to buy a 24GB RAM Mac Mini but the seller rescinded due to the fact their child really needs it for college, and I agreed, so I am about to move forward with a Macbook Pro M2~ with 64GB Ram.

So my question to you is, is vibe coding tools viable?

Thanks

 

[casperes1996]

It really depends on the scope, ease of validating correctness, etc.

If you can somewhat easily test and validate correctness and the scope isn't huge, you can definitely vibe it. The better you are at documenting what you want the better the outcome. As scope increases and validation becomes harder, more software engineering skill is needed on your part to steer the tooling and correct it, and at some point it's still easier to write it all your self. For skilled programmers, generating code was never the hard part. These tools are definitely fast and decent at generating code, but for something that needs to scale they may very well architect you into a ditch. But if you never need much scalability they can get you results

 

[haralds]

Absolutely! I have built several consumer apps for personal use and ported an SDK that manages real-time calculations of physiological data. I am now working on a commercial product for cross-platform deployment. They are native on macOS, iOS/iPadOS (Swift), Android (Kotlin), Web & Chromebook PWA (TypeScript), and Windows (C#). The capabilities have improved significantly over the last six months. With my training and careful prompting, Claude gets most things right out of the shoot or with a few corrections. In the last couple of months, UX validation has much improved.
I develop on macOS using Claude Code hosted in VS Code with a few additional tools, plus Xcode. I call it meta coding; there is nothing "vibed." Everything I do is carefully controlled and specified, from architecture to code layout, and code quality.
Scaffolding is critical. CLAUDE.md and AGENTS.md define the project framework and choices. They are project-specific.
I prompt in an MD file, PROMPT.md. I do not use verbal interactions. I like to think elements through. Claude tracks its responses and decisions in CHAT.md. Architectural choices and issues are tracked in additional documents along the way.
A repository is absolutely required. I use Bitbucket with work projects, but also have Gitea as a Docker instance on my Synology NAS in my home office.
Numerical computing tasks are a cinch if you have qualified input and output data. You can have the AI work to completion.
Keeping the documents and project well organized means that Claude Code can learn your style and preferences, and get smarter about the project along the way. I tend to work on three different Macs; the repo links the project and Claude work history.
I used to prefer Cursor, but we have an Anthropic Team membership at work, and I have never burnt through my token limits the way I did with Cursor. Opus 4.9 and 5, and now Fable 5 are great with Swift. C# and Windows UI are more of a challenge. I model everything on macOS and iOS.

 

[haralds]

One other note. Do not have two agents run on the same project tree. Even with care, it often leads to problems and conflicts.

 

[poorcody]

Yeah definitely doable. I'm still amazed at it sometimes. I would say just be sure to do a lot of testing of the code it generates, and it's worth reviewing the code architecture regularly, to make sure the AI isn't just building a spaghetti mess. Tell it not just what you want but how you want it to organize things. The more detail you tell it the better -- don't be shy. But I think you'll find it quite good for your needs.

 

[poorcody]

Absolutely! I have built several consumer apps for personal use and ported an SDK that manages real-time calculations of physiological data. I am now working on a commercial product for cross-platform deployment. They are native on macOS, iOS/iPadOS (Swift), Android (Kotlin), Web & Chromebook PWA (TypeScript), and Windows (C#).

So you are using different languages and libraries for each platform? I'm curious how you approach that with AI. Do you design on one platform, and ask the AI to produce the same thing for other's?

I can't tell you how many cross-platform frameworks I've used over the decades, and nothing ever quite worked well enough. Using AI as a tool to do native everywhere is intriguing. Curious how you approach it exactly... ?