@tjrobertson52

The way people prompt AI models has changed dramatically in the last six months. So if you're still prompting like it's 2025, I wanna catch you up to speed. I'm gonna give you three things that have changed, starting from the simplest to the most advanced, and I guarantee if you implement these changes, you're going to see a dramatic improvement in your output.

First is we don't write prompt templates anymore. We create skills. What are skills?

They're just formalized prompts. Templates create these ones, and they can live inside of Claude or in Codex or on GitHub. It doesn't matter, but this is how it's done now.

It's the easiest way to reuse instructions and share them with each other. Second is that for most tasks, you don't want to write out rigid step-by-step instructions. Instead, any skills or design specs that you're writing for Claude should be written the same way you would write those documents for a senior-level member of your team.

In other words, focus on the goal. What does success look like? Take time to describe it clearly and in detail, and then provide all the necessary context that Claude or whatever model you're using is going to need to accomplish that goal.

But as these models get more intelligent and more capable, giving them specific rigid steps actually hinders them. As long as they clearly understand the goal, you're best off letting them come up with the plan and steps to implement that goal. And the smarter the models get, the more true this is gonna be.

So if you don't want to be rewriting your skills every time a new model comes out, this is the way. And the third and most advanced tip is for big, complex projects, don't write prompts. Write loops.

If you're creating software or a website or a long analysis or a very complex plan, anything

big, complex projects, don't write prompts. Write loops. If you're creating software or a website or a long analysis or a very complex plan, anything where even Claude is going to need to work on this for hours or maybe days, you don't want to be prompting Claude and then waiting for the output, evaluating it, and then prompting it again.

Instead, you want to write a very robust specification document, and of course Claude can help you create that document. And then in Claude Code, ask Claude to run a loop against that specification. It should continue prompting itself until that specification is met.

With that loop command, Claude can run for many hours straight, and Codex has the same feature.