Everything changes every two months. New models drop. New harnesses. New agent frameworks. The benchmarks shuffle. Twitter loses its mind. And you’re supposed to have opinions about all of it?
Yes. Absolutely yes.
Why bother having opinions?
You have opinions about languages. Python for scripts, C++ for performance, TypeScript for the web (unless you’re DHH, who dropped TypeScript from Turbo 8 because it “pollutes the code with type gymnastics.” Simpler times. That was two years ago. How is that possible). You didn’t get those opinions from Twitter threads. You got them from building things, hitting walls, learning what breaks where.
Models are a layer of abstraction on top of languages now. And harnesses on top of models. And agent frameworks on top of harnesses. You need opinions on all of it. You should know when to reach for Opus vs Sonnet. When to use Codex. When Claude isn’t the right tool at all. When Pi might actually be what you need for brainstorming. These aren’t hypotheticals anymore. You’re using these things all day. They’re as fundamental to your workflow as your language choices.
So why do so many engineers just… not have opinions here?
Twitter is not the way
How do you form these opinions? Twitter? Not really. Twitter is the news now. People get paid to post. Hot takes are profitable, and “X model is cooked” gets more engagement than “here’s a nuanced comparison of context window handling across three similar tasks.”
Twitter is fine for finding out what’s going on. What just dropped, what people are excited about, what’s worth looking into. But it won’t tell you what’s useful or what actually works in the specific areas you’re an expert in. That you have to figure out yourself.
Use them for things that matter
So find the things people are talking about and use them. Use them for things that matter to you. Find things that matter to you.
If you use these tools for your work, and you are not deeply invested in your work, you will learn nothing. Who cares if the output is subpar if you still ship the feature? This is still useful! But you aren’t investing in your growth as an engineer. And that matters more right now than it’s ever mattered.
When the code ships regardless of quality, the feedback loop breaks. You don’t learn which model handles complex refactors better. You don’t notice which one loses context after 20 turns. You don’t know which model has better ergonomics when you’re in the loop and which one is better to let spin overnight on a prd.json.
You need stakes. Personal stakes. Code you care about. Problems you want solved well, not just solved.
Write a coding agent
You should write a coding agent. There’s really so little to these things and you’re going to use them all day. They made you write a compiler in college for CS. It’s a matter of time before writing a coding agent is one of those shared traumas people talk about at meetups.
Build something small. Hook up the APIs. Watch how different models handle the same prompt. See where they fall apart. See where they surprise you.
The gap between “uses AI tools” and “understands AI tools” is widening fast. You can’t coast for three months anymore without it showing. The learning compound interest works both ways.
The opinions themselves are temporary
Here’s the thing. Your opinions will be wrong in two months. The model you swear by will get dethroned. The framework you dismissed will become essential. That’s fine. That’s the point.
Having opinions isn’t about being right forever. It’s about having a framework for evaluation. A lens. When the next model drops, you know what questions to ask. You know what tasks to test. You know what matters to you.
When the next model drops, you already know what to test. You’re not starting from scratch every time, waiting for someone else to tell you if it’s good.
So what are my opinions?
Right now? Opus for almost everything. I don’t even use Sonnet. Sometimes I grab Codex-5.2-max when I want a critical eye on something that touches a lot of pieces.
But ask me again in two months. The specifics change. The practice of having opinions doesn’t.
What are yours?