🤓 The New Rules of Prompting (and a Prompt to Rebuild Yours)
A practical guide to getting the best results from GPT-5.6 and the next models from OpenAI and Anthropic—with a copy-and-paste prompt to update your existing prompts and project instructions.
For the first few years after ChatGPT launched, getting better results from AI meant becoming good at prompting.
And I got really good at it.
I wrote longer, more detailed prompts.
I gave models elaborate role assignments and broke complex tasks into step-by-step processes.
I repeated important instructions and asked AI to share its thought process.
I even told the models to take a deep breath before getting started (if you know, you know).
Those techniques worked.
The models needed more guidance, and learning how to provide it led to noticeably better results.
But over the last year, I’ve had to unlearn much of that.
The better these models get, the less we need to tell them how to do the work.
The newest models, including OpenAI’s GPT-5.6, can understand more, plan better, and work out far more of the process on their own.
They’ve also become much better at following instructions and making sense of large, complicated bodies of context.
They need less direction about the how—and much better information about what you need, why you need it, and what a strong result should accomplish.
Which means our approaches and habits have to evolve too.
Mine have had to. Repeatedly. And sometimes reluctantly.
The deep breath was the first to go.
I’ve been doing this long enough to watch my own advice expire.
This Sunday marks two years since I sent the first edition of this newsletter.
Two years and 81 editions, which in AI years is roughly a century.
All of it has been free, which means my paid subscribers have been paying for one reason: to support me and this work. 💙
For this anniversary, I wanted to thank them with something more than a thank-you note.
Which is why this week, you’re getting a very special edition.
I’ve turned what I’ve learned over the past year, the stuff I’d normally cover in a workshop, into a practical guide for getting better results from this new generation of models.
The guidance and principles apply to the most capable models available today, including Anthropic’s Fable.
They should hold for the next wave from OpenAI and Anthropic too, because they reflect where these systems are heading.
The guide also includes some examples as well as a prompt you can cut and paste into ChatGPT or Claude to rebuild your existing prompts and project instructions for these newer models.
Most of my time these past months has gone into figuring out the agentic tools—Codex, Claude Cowork, and now ChatGPT Work—the ones that take actions instead of just giving you answers.
They run on the same fundamentals as everything in this guide, but with much higher stakes and a steeper learning curve.
Learning happens in layers, and this is the layer worth getting right first.
Use this guide to update your approach and systems, so you’re ready for the agents.
They’re already here.
Paid subscribers can continue reading the full guide below.
If you’re a free subscriber, you can upgrade below to access the guide and get the copy-and-paste prompt.
And I’ll see you all in a few weeks for the next regular edition of the newsletter.