Do you still write code “from scratch” or mostly remix and adapt now?
by•
I’ve noticed that my workflow has changed completely over the last year. I rarely start a new project with a blank file anymore. Instead, I pick a template, reuse snippets, or let an AI helper suggest the structure and then I just vibe my way through the build.
It’s faster, but sometimes I miss the old “blank screen energy,” when every line felt handcrafted.
I’m curious how others here approach it:
– Do you still prefer to build from scratch?
– Or has remixing and fast iteration become your new normal?
– Do you think this shift is making coding more creative, or less intentional?
I’d love to hear how your process has evolved and what “vibe coding” means for you right now
231 views
Replies
Triforce Todos
AI has made me focus more on ideas than syntax. So creativity just moved one level higher.
@abod_rehman That’s a great way to put it .it really feels like syntax is becoming invisible, and the creative layer is what matters most now.
Almost like we’re designing logic more than writing code
@abod_rehman Precisely! When I get an idea, I usually run it through deep research and get to building with the system architecture that I built with AI.
Inbox Agents - Intelligent Unified Inbox
@cam_mcmaster1 That’s a really solid mindset , I like that distinction between “understanding first” and “then letting AI take over.”
I’ve found the same thing: if I skip that early “manual build” phase, I end up not really grasping what’s happening under the hood later.
Reading every line sounds slow, but it’s probably the only way to stay sharp while still taking advantage of AI speed.
Remixing has become my norm. I look at it like LEGO. I’m still building cool stuff, just with smarter blocks. @minach
@orthopedic_surgeon1 I love that analogy , “smarter LEGO blocks” sums it up perfectly.
It’s wild how remixing doesn’t really feel like cutting corners anymore, it’s just a new kind of creativity.
Mindstimer
If we rely to "remixes", exactly the point in history when they stopped to call it "recordings" and started to call it "tracks".
Punk rock era of IT is over, hip-hop is coming
@butaji That’s such a great metaphor . I love the “punk rock vs hip-hop” take.
It really feels like we’re sampling, remixing, and layering code now instead of just writing it raw.
for me I don't build anything from scratch. I start with analyzing the project and setting up requirements for the first main files and main functionlity, then I discuss things with gpt or claude and argue with them and do some back and forth.
and once things feel clear, I make it build the files, with the things we agreed on. And I don't really make it doing full project at once even if it's not a big project. I just like to stay connected to the structure of my project.
Then and this is crucial, I tell it give a markdown .md document, containing information about the project idea and what we are trying to do, and down it explain shortly what we have built so for. and what will be the structure of the project that we already agreed on. And things work very well.
@sharif_hasan4 That’s a really thoughtful workflow ,I like how you treat GPT and Claude almost like design partners during the planning phase.
Having them generate the structure and even the .md docs sounds like a great way to keep everything aligned.
Do you find that this setup helps you stay more consistent across projects, or does it change depending on the scope?
@minach it helps so much, I don't really change this setup that much from project to another.
Snowglobe
If I’ve already built a similar project from scratch before, I’ll totally “vibe code” it
(Typically by getting Blocks to create an Epic on Linear and break it down into manageable sub-issues)
But if I haven’t, I think everyone should go through the exercise of writing it themselves at least once because you learn a lot that way
For me, anything I’ve done many times and consider “easy” (from a Staff Engineer perspective) is perfect for agents.
These days my workflow looks like this:
Easy or low-medium tasks → background agents (I use Blocks with Claude Code, Codex and Gemini CLI) handle them end to end with no intervention
High-Medium or hard tasks → I pair with an agent in the CLI or IDE, guide it, and write the opinionated or nuanced parts myself
It has been a good balance between speed and control
@alejandroesquivel That’s such a balanced approach , I like how you split it by difficulty level.
I’ve been experimenting with something similar lately: letting agents handle the repetitive setup and boilerplate while keeping the “thinking” parts for myself.
It feels like the sweet spot between flow and control.
Curious , have you noticed any point where the agents start making too many assumptions, or do you usually step in early to correct their direction?
Snowglobe
@minach For the tasks I let agents handle in autopilot, low effort tasks I feel can be one shotted with the latest models, for medium sized tasks I usually groom a Linear/JIRA ticket along with an agent which narrows down the directions it could go
Nice topic. To me, it depends :)
I could create from scratch and an empty project a proof of concept of a simple mobile app in Android Studio by connecting it to Gemini API, and prepared to launch it here too in just some hours of work having never worked professionally with flutter.
I spent some dollars for some hours with it, but the result was amazing and to start it was great.
I couldn't find the same functionality using Rider with .NET projects, for example. So for an API, always to be used as a proof of concept, I could ask ChatGPT and get a solution ready to start. For bigger projects I have a .NET template that I created in the past, that works well and I keep using as a base for it.
Generally, I am still convinced that you always need to know your language, and the best practices to follow. I had to guide the AI continuously asking for improving the classes, refactor them to not duplicate the code (and improve the speed of its code generation), etc... I aim to a good start with a minimum idea of architecture and chose the right approach in base of the type of project I need to start.
I've been a software developer for 20+ years, worked in over a dozen companies, mostly early-stage startups. I've built systems with C++, Java, Python, Scala, Node.js, Typescript, Ruby, Erlang....
Over the past 2+ years I have rarely coded a product feature by hand-writing code in anything other than English!
Prompting, first via browser chats, then via more and more sophisticated IDE plug-ins, OpenAI's Codex , Junie, Copilot, Cursor, etc..
Most of the code I write these was edits, rewrites, polishes, and fixes over what AI agents generated.
This has sped up product feature development by more than 10x...
It has also allowed me to focus my attention on higher level things, such as overall systems architecture, product features, cloud infrastructure, etc..
Coding by hand is definitely going to be a thing of the past, regardless of how much people (including myself) will miss it.. The AI is just getting much better (and faster) than humans at this.
I wrote a post recently about my current approach when I start a project from scratch:
https://olegkozlov.dev/posts/agent-driven-development-1
It was about 2 years ago I started to use ChatGPT for my coding tasks and at that time I only asked questions regarding features or algorithms that I hardly understood the whole structure or something. I'm using Claude AI more than GPT for coding recently and still trying to hand code as much as I know and can do myself, but I think using AI tool to write repeated code much faster and learn coding methods I'm not familiar is very useful, efficient. To be honest, I even believe that it is actually almost one of the mandatory part that we developers need to accept to do more creative and innovative programming and designing whatever the type of software or application we develop like game, mobile app, web template. We can spend more time on the things that AI can't do itself (at least for now) such as beautiful front page design, a backend feature specialized only for a specific business or user, and so on. Of course it is very important to read and check every code line written by the AI to make sure there is no mistake or bug, but it's still very helpful and makes my work much better.
I dont think theres ever been a case of building much on the web that hasn't been copied/pasted/ai'd or something similar -- we look at art of the ages, everything informs the future and artists and designers, musicians and architects have always "copied" and been inspired by previous or contemporaries.
Ai has allowed automation and speed to a cumbersome process - and everyone uses it slightly differently -- to their advantage.
vibe coding Ergo -Keys to the kingdom of the web, has allowed me and many others to - like already said here, to focus on concepts and ideas business or otherwise that were previously too difficult and thus expensive to implement though manual coding..