Comment by ryanjshaw
18 hours ago
Until a year ago I believed as the author did. Then LLMs got to the point where they sit in meetings like I do, make notes like I do, have a memory like I do, and their context window is expanding.
Only issue I saw after a month of building something complex from scratch with Opus 4.6 is poor adherence to high-level design principles and consistency. This can be solved with expert guardrails, I believe.
It won’t be long before AI employees are going to join daily standup and deliver work alongside the team with other users in the org not even realizing or caring that it’s an AI “staff member”.
It won’t be much longer after that when they will start to tech lead those same teams.
The closer you get to releasing software, the less useful LLMs become. They tend to go into loops of 'Fixed it!' without having fixed anything.
In my opinion, attempting to hold the hand of the LLM via prompts in English for the 'last mile' to production ready code runs into the fundamental problem of ambiguity of natural languages.
From my experience, those developers that believe LLMs are good enough for production are either building systems that are not critical (e.g. 80% is correct enough), or they do not have the experience to be able to detect how LLM generated code would fail in production beyond the 'happy path'.
The amount of "apps" I've had dumped on my team that are everything from un-releasable to deployed on some random shit-cloud we haven't approved (vercel comes up a lot). If you needed hand holding to release things or had to throw software over the fence to others to "productionise" etc then you probably don't know what you're talking about.
This is not my experience with claude code. It does forget big picture things but if you scope your changes well it’s fine.
I would estimate that out of every 200 lines of code that Claude Code produces, I notice at least 1 issue that would cause severe problems in production.
In my opinion these discussions should include MREs (minimal reproducible examples) in the form of prompts to ground the discussion.
For example, take this prompt and put it into Claude Code, can you see the problematic ways it is handling transactions?
---
The invoicing system is being merged into the core system that uses Postgres as its database. The core system has a table for users with columns user_id, username, creation_date . The invoicing data is available in a json file with columns user_id, invoice_id, amount, description.
The data is too big to fit in memory.
Your role is to create a Python program that creates a table for the invoices in Postgres and then inserts the data from the json file. Users will be accessing the system while the invoices are being inserted.
---
7 replies →
Yes, but knowing how to scope your changes requires a lot of expertise.
After 2 years of using all of these tools (Claude C, Gemini cli, opencode with all models available) I can tell you it is a huge enabler, but you have to provide these "expert guardrails" by monitoring every single deliverable.
For someone who is able to design an end to end system by themselves these tools offer a big time saving, but they come with dangers too.
Yesterday I had a mid dev in my team proudly present a Web tool he "wrote" in python (to be run on local host) that runs kubectl in the background and presents things like versions of images running in various namespaces etc. It looked very slick, I can already imagine the product managers asking for it to be put on the network.
So what's the problem? For one, no threading whatsoever, no auth, all queries run in a single thread and on and on. A maintenance nightmare waiting to happen. That is a risk of a person that knows something, but not enough building tools by themselves.
Yup. I’m not expert so maybe I’m completely off base, but if I were OpenAI or Anthropic I’d likely just hire 1000 highly skilled engineers across multiple disciplines, tell them to build something in their domain of expertise, then critique the model’s output, iteratively work on guardrails for a month or two until the model one-shots the problem, and package that into the new release.
That's exactly what they are doing via dataannotation.tech and other services.
Any comments on how the copyright issues are handled in corporate settings? I mean both in terms of staying clear of lawsuit+ ensuring what we produce remains safe from copying
I can take a verbal description from a meeting with five to ten people and put together something they can interact with in two weeks. That is a lot slower than Claude Code! Yet everywhere I’ve worked, this is more than fast enough.
Over two more weeks I can work with those same five to ten people (who often disagree or have different goals) and get a first draft of a feature or small, targeted product together. In those latter two weeks, writing code isn’t what takes time; working through what people think they mean verses what they are actually saying, mediating one group of them to another when they disagree (or mostly agree) is the work. And then, after that, we introduce a customer. Along the way I learn to become something of an expert in whatever the thing is and continue to grow the product, handing chunks of responsibility to other developers at which point it turns into a real thing.
I work with AI tooling and leverage AI as part of products, where it makes sense. There are parts of this cycle where it is helpful and time saving, but it certainly can’t replace me. It can speed up coding in the first version but, today, I end up going back and rewriting chunks and, so far, that eats up the wins. The middle bit it clearly can’t do, and even at the end when changes are more directed it tends toward weirdly complicated solutions that aren’t really practical.
> poor adherence to high-level design principles and consistency. This can be solved with expert guardrails, I believe.
That’s a bit… handwavy…!
I've been hearing this for several years. How much longer is "it won't be long"?
I've heard the same "it won't be long" from UML and 4GL - until the industry finally gave up. Both of those are still used a lot in industry and they do well in their place, but nobody pretends they will ever be everything to everyone anymore.