@unclebobmartin I think that has been the hardest part of AI adoption at organizations. The business is sold an idea without being educated on how non-deterministic applications will be have. So when we give the demo they expect the exact same output every time.
@lochan_twt I will keep the code open in VS Code and use Github Copilot to refactor stuff that gets too crazy. Having good instructions will help it follow project design patterns too.
@paradite_ I have had the same experience so far. Codex has taken longer and done weird hacks instead of solving the real problem compared to Claude code.
If you haven't used Claude Code cli or Github Copilot(GHC) cli, it is worth a shot.
The last few days I am starting to understand when they say 10x engineer.
I used to use GHC in VS code only and it was good. I did alright. Now I am throttled by the amount of context in my own head.
Start by creating a UI, API, and database agents that know how to do their job and understand your expectations well. Then you can have Claude Code doing more of the complex work(create an ai assistant chat bot that edits text live) and GHC building more crud (a task list or managing members) they will both be working through their plan and finally I use GHC in VS code to go back and do any refactoring if the AI got a little to verbose.
Most of it is sonnet 4.5 but will use gpt4.1 for refactor.
I’m reading Thinking in Systems, and the book shares a simple story about two watchmakers.
Both build the same watches.
Both face constant interruptions.
Yet one stays poor while the other becomes wealthy.
The first watchmaker tries to assemble the entire watch at once. When a customer interrupts him, the watch falls apart and he has to start over.
The second builds watches in stages. Small, stable subassemblies are completed first, then combined. When he’s interrupted, only the last subassembly is affected.
This allows him to serve more customers and finish more watches.
It got me thinking about all the things I’m trying to “build all at once”, when they could be broken into stable stages and built more effectively.
Anyone else impressed when getting AI to create HTML mocks it genereates the CSS first?
I have a hard enough time adding CSS when I know where it is. I can't imagine predicting these styles will work, let alone make the app look good.
To the vibe coders out there and anyone else using React Server Side components - please upgrade your React/NextJS/React Router packages. There is a critical security vulnerability allowing for unauthenticated remote code execution.
This allows attackers to send unauthenticated malicious scripts to your server and do bad things.
If you want to know more about which packages and the vulnerability, I linked the React team's blog linked below.
I’m learning growth can feel like two opposite emotions at once.
On one hand, there’s excitement from the momentum, progress, and new skills clicking into place.
On the other hand, there’s the frustration of dropping things, forgetting things, and feeling like I should be keeping up better.
What I realized I need to work on:
- Celebrating the excitement and the wins
- Giving myself grace for the things I’m still figuring out
If you are feeling both the momentum and the mess, just know, you are growing.
Today I learned a simple framework that can help make idea clearer:
Character → Struggle → Resolution
Example:
Character: a shipper you're trying to help
Struggle: the chaos they’re stuck in (late deliveries, bad visibility, angry customers)
Resolution: a life of on time deliveries, clear visibility, and fewer dumpster fires after you help out.
Do y'all use stories often to express ideas?
If you are having GitHub Copilot in Agent mode doing your tasks, try adding the prompt below to the rest of your prompt. It has been an amazing addition for me.
P.S. Add it as a prompt file or to your copilot instructions to avoid having to type it every time.
---- Add below to your prompt ---
- Break the request into multiple tasks.
- Work through each task one by one. Only moving on once the current task is completed.
- Ask for any additional information before beginning if the request is ambiguous or lacks sufficient detail.
Me: Hey Claude, I am a perfectionist and don't like to make uncomfortable calls so I am going to word vomit to you instead...
Claude: I'm Drawing a Line
I will not answer any more questions about this idea until you make the calls.
@DataChaz@n8n_io Thanks for the share, I have just started looking into observability for enterprises and was wondering how that might apply to nocode.
Would you like your own personal LLM?
Most LLM models are able to be finetuned for a pretty reasonable price nowadays. I created a Derek-LLM based on the gpt4.1-mini model and it is able to create LinkedIn posts and Xeets with a very simple prompt.
I am not sold on the content it is creating yet though... or am I?