Dean

Senior Machine Learning Engineer in FinTech | Python, K8s, Spark, Distributed Systems | MSCS @GeorgiaTech '23 🐝 | Opinions are my own

[Dean @CodeWithDean]

I spent weeks confused by Claude Code's feature ecosystem.

Skills? Subagents? Plugins? Commands? Agent Teams?

The mental model that finally made it click:

Skills are knowledge.
Subagents and Agent Teams are workers.
Plugins are how you package and share all of it.

Wrote the article I wish I had from the start.

https://t.co/79YLCSbMrP

[Dean @CodeWithDean]

@trq212 @johnerik Do you recommend transforming this prompt into a Claude skill, and manually invoke after work is done?

[Dean @CodeWithDean]

@theo Pretty sure the duplicate window you have up is meant to have same typing in both input boxes as intentional feature

[Dean @CodeWithDean]

@HankFrank Same thing happened to me. Did 2 years. Not renewing

[Dean @CodeWithDean]

Tokenmaxxing

[Dean @CodeWithDean]

What do you mean can't be trusted with prod?

Are you not looking at the code the AI is producing? No agent should be trusted with prod without human in the loop.

The top LLMs are all good enough at this point to do anything you tell it. If it's not producing results you want, you're not planning enough or being detailed enough. Your dislike for Anthropic doesn't seem authentic tbh. Almost as if you're chasing an open ai sponsor 💰 😭

[Dean @CodeWithDean]

@theo This comparison doesn't tell the whole story. Does your app offer the same exact user experience and features?

[Dean @CodeWithDean]

@theo This is likely the result of bad planning and not giving it a test harness while it worked on your feature

[Dean @CodeWithDean]

https://t.co/u4DORLvQsU

[Dean @CodeWithDean]

Most people still use AI the way they did in the ChatGPT era.

Type a question. Get an answer. Copy-paste it somewhere. That's how it worked when ChatGPT came out in 2022, and most people haven't updated their mental model since.

Starting in 2024, we entered the agentic AI era. Tools stopped just answering questions and started doing work. Claude Code is at the center of that shift. It's an agent that reads your files, runs commands, makes decisions, and operates inside your actual codebase.

Even out of the box, it's powerful. @bcherny, the person who built it, has said his own setup is "surprisingly vanilla."

But there's a gap between "it works" and "it works the way I want."

Without the right habits, you end up generating "AI slop" you'll eventually rewrite anyway. I learned that the hard way.

So I built a simple two-part framework: compounding habits you set up once that make every session smarter, and session habits that keep each task on track.

If you want to get reliable, quality output from Claude Code, this one is for you.

Link in comments. 👇

[Dean @CodeWithDean]

@daviddorg @superpower @ouraring can't be possible. how will you use twitter to provide updates? how will you file for your taxes or view your blood work results without a computer?

Also, what kind of phone will you be using that does not have a screen?

[Dean @CodeWithDean]

@bcherny @cvaartjes More powerful in what sense?

[Dean @CodeWithDean]

I often catch myself typing full English sentences into a terminal, forgetting I'm not talking to an LLM

[Dean @CodeWithDean]

@bcherny thanks for making Claude code. It's by far my favorite agentic tool and my whole team prefers it over others.

Have a highly useful feature request for Agent Skills if you happen to see this. We have shared skills across our organization. Many skills are tied to usage of specific internal libraries on GitHub. If these libraries update, the skills may become stale (outdated method interface, etc). Can your team incorporate a mechanism natively to keep skills updated (e.g. pull latest git tag and read latest code) if skill is falling behind?

[Dean @CodeWithDean]

Hot take: the era of dual monitors and ultrawide setups is ending. Agentic AI handles the multitasking now, so what's the point of having massive screens anymore?

You just need one clean screen to supervise it. Get yourself a single 27 inch monitor and call it a day.

[Dean @CodeWithDean]

Claude Code is coming for your job