I've watched a few with Jack Roberts recently, ( https://www.youtube.com/@Itssssss_Jack ) - lately he's done a few showing use of antigravity -

not sure these would qualify for 'good resources' - that really depends on many factors. However at least one or two of his showing 'how to use' does check the boxes of your 1,2, and 3 - surely not a complete "here is everything you can do" - but not sure such a thing exists.

maybe good tips: https://www.youtube.com/watch?v=zxMjOqM7DFs&t=541s

I liked info from the talks here: https://www.youtube.com/watch?v=hOqgFNlbrYE

I have some open tabs for going through the learning the openai codex has online - I poked into a couple and they seem very good.

Running openai codex inside vscode right now and claude code via web -

will be setting up a new computer do run these things, not sure I trust the power they have to run on main box.

Anthropic's prompt engineering docs (docs.claude.com) are secretly the best AI coding guide teaches you how to structure requests that actually work. For tactics: search "AI coding workflows" on HN/Reddit, filter for comments with war stories not marketing. The people complaining about what broke have better insights than the people hyping what shipped.

Thanks, this sounds interesting!

using them. there really is no other way.

It's like asking "how do I use my text editor".

Almost everything you read online will be out of date, and the person working on it won't work the same way you do.

Some people love it, some people hate it.

And unless you've got your own experience, it's hard applying other people's experience to your experience.

Especially for something like Claude Code where you're just prompting and getting results back.

I mean, half the time I use VS Code, half the time I use a terminal window.

You're going to get a lot of conflicting advice because everyone's environment is different and they work on different sorts of code bases.

But I'll give you mine.

I'm a Claude Max Pro 200 subscriber.

I sit with the Opus chat-bot have a discussion and come up with a spec. We turn this into a zip full of documents and upload it to a GitHub Repo.

e.g. https://github.com/lawless-m/BattleForMoscow

I'll then get Claude Code for the Web to go to that repo unzip the zip and read the documents. It will make a first pass at the entire codebase.

I'll merge that into main and create another Claude Code for the Web Opus session with any ideas I've had in the meantime - which will usually be a few.

Then I clone it to a local machine and get Claude Code Opus to try and get it to work. And I'll prompt it from there until it works. If it's a Linux program, that'll be in a terminal window. If it's Windows, I'll use VS Code because it's a better terminal in VS Code than it is in a terminal window on Windows.

That's a general workflow. Sometimes I won't use GitHub at all. Sometimes a PXE boot an entire Linux machine and give it that with admin privs.

And sometimes I just tell it to use sudo as my own account. On my router for instance, if we want to do things with the firewall.

Why do you go through all the trouble of uploading to Claude code for web only to download it back and run it in Claude code terminal? Are there different rate limits for each endpoint so that’s why you do it? Why not just work entirely using Claude code terminal locally?

I do most of my initial work in Claude Chat, on my phone. I have my best ideas when I'm away from my desk. Originally, Claude for the Web was only on the iPhone not Android - but it is now.

My iPhone 11 won't let you download certain files, and file handling on the iPhone is awful.

I guess I've just got used to the flow. I don't always do it like that, it was just one example.