Wow, this thread got bleak. This is like one of those dystopian television shows that you occasionally tune into that remind you: ‘no matter how bad your life is, it could get so much worse…’
It is truly ushering in a boring dystopia that nobody would actually watch every episode of.
Speaking of though. I asked Claude to imitate Marvin from hitchhikers guide in responses to me and it’s made it a lot more tolerable now that Claude sounds even more depressed and miserable to be there than I am.
I am not sure what tasks you guys are doing, but I am still finding that AI cannot develop code for real-world application code-bases. It can create great, syntactically-correct standalone features; but integrating with an existing code base and integrating with existing features is not in its wheelhouse.
I also find that it is more time-consuming to create a bunch of .md prompt-documents than it is to just develop the feature myself. You all must be doing things I am not.
I mostly use it to augment my development. Stuff like “I have an array1 of objectA and an array2 of objectB. Create an array3 of fields that are common to both objects in array1 and array2.”. But eve that kind of stuff saves very little time between just doing it myself.
It sounds like people were saying they don’t have a choice, I don’t think anyone was saying it’s more efficient or had better output. And they are probably also using different LLMs.
Correct. It’s not optional. I will also say that provided you don’t care about generating tech debt, and upper management simply doesn’t care about that, then if you use Clause Opus 4.6 you can honestly have it code just about anything. I created a large scale vpn service in a week. I recently also had it take an existing baseline web app that meets company standards and create a new one that matches those standards.
However you have to understand the codebase and be able to truly review the code. It’s like asking my kids to clean their bedroom. They are gonna say they are done but left crap everywhere and I’m gonna have to come and point out all that crap.
It’s honestly still faster than hand coding, but only if you have the right model and know how to effectively use it (spec coding and not vibe coding).
I was simply reacting to everyone describing how AI was making their jobs more efficient and more productive. My experience is just different. I am using it; just very selectively. Its use is definitely a trade off between describing complex features and business-logic for prompting AI; or just doing it myself integrating with the architectural and technical framework of the application that I am already familiar with.
I think a lot of it is the model you are given access to. Copilot vs clause opus… Is like the difference between some outrageously expensive sports car and a Ford Fiesta.
I use Claude Code at work, and set up rules for Model Routing so that all my requests go through Haiku, and then it spins up subagents with the appropriate model for the task. I’m doing general web dev at an advertising agency, so it’s ended up saving me tons of time. I have very strict rules on structure, and end up having to review code that’s mostly pretty tight for our needs.
Currently I am using Claude Opus.
Copilot allows you to use any number of models. Each cost different $ so we are told to use them in just the way you describe.
If you do research into spec coding and how to set yourself up I think you will be surprised at what it can accomplish. It’s also frustrating and significantly less fulfilling than doing it yourself though so if nobody is forcing you, you might choose to maintain your sanity instead.
My entire job has become “human guardrail,” making sure we don’t acquire tech debt. They just LOVE forcing us to use AI.
Does spec-coding work for an existing code-base? That is where I see the problem. It does not understand how corporate component libraries work and just starts writing its own HTML. So, I then have to go back and re-implemnt all the UI code using our company component-library and correct CSS.
Yes. Spec coding is a system ( there are formal systems but you can kind of roll your own) where you specify in painful detail what output you want and your input.
An easy way to get started would be like hey Claude please analyze this existing codebase. Take note of all these important corporate components. Generate a document outlining all important elements. Then you manually review the document and make changes. Then you can say hey Claude I need to create a new feature for this site, review this staging document and generate a requirements document that meets these user stories and these feature details. Review that manually to make sure it covers everything. Hey Claude generate a design document. Review that mermaid design doc. Hey Claude generate a task list of individual tasks to accomplish this design meeting the specs. Review that. Hey Claude action these tasks. Review all code output and point out any mistakes.
Things like Code Sight, RTK, and the Superpowers plugin, also help things clean, and use less tokens.
I have found it takes less time to just do it myself.
But yes; I am in the process of creating Storybook documentation that also creates documentation that is geared towards AI use for our component-library so that teams in the company can do exactly what you are describing.
I am using it a lot lately. My ex wife defaulted on the consent judgement and i am taking her back to court and it’s helpful to discuss things with Gemini before contacting a lawyer. I am meeting with an attorney tmrw and I’m ridiculously prepared










