GPT-6 Astra: new models need new prompts

Last Thursday I started using GPT-6 Astra, OpenAI's newest model. I have used every major model for the past three years, and since Claude Opus 4.5 came out last year I had not felt a jump like this one. It feels like a different level of intelligence, not a better version of the same thing.

It also changed how I prompt. I started noticing that my prompts were sometimes not helping the model do its job. They were getting in its way. So I went reading, and this turns out to be the general view among people working with these new models: skills and very detailed prompts are now hindering them more than helping them. OpenAI says the same in its own guide for Astra. Guidance that used to help can now hold the model back.

I share this because the landscape changes fast. It sounds counterintuitive, and it goes against a lot of what we have all been learning. But it is where the models are right now.

The difference between briefing a senior partner and a junior associate

A way to think about it: the difference between briefing a junior associate and briefing a senior partner. The junior gets a checklist and does exactly what it says. The partner gets the goal, the context, and what good looks like, and works out the rest. Prompting a strong model the junior way wastes the partner.

What does that look like in a prompt?

Over-prescriptive, the way many of us learned to write:

You are an expert pricing consultant. Step 1: list five pricing models. Step 2: score each one on cost, simplicity and perceived value. Step 3: build a comparison table. Step 4: recommend one. Step 5: write a one-paragraph summary. Do not skip steps. Do not add anything I did not ask for. Follow these rules exactly.

Right-sized for a strong model:

I am launching a small live workshop for past coaching clients. Here is who they are, here is what I charge for my other work, and here is what I want to avoid. Tell me how you would price it and why, then argue against your own answer.

The first prompt gets you a tidy table. The second gets you a thinking partner.

So are skills and precise prompts useless now?

"You have been teaching us skills and precise prompting for months. Is all of that useless now?"

Absolutely not. There are still jobs that need a procedure and a template: the weekly report, the standard summary, anything that has to come out the same way every time. That is what skills are for. What changes is where you use them.

People who watch their budget are also learning which model does which job. In my experience so far, the big senior models like GPT-6 Astra and Claude Fable 5.1 do their best work with less instruction and more freedom. That is where your thinking, your decisions and your new projects go. Smaller models like Sonnet and Haiku, and sometimes even Opus, do better with a tighter structure and clearer steps. That is where your procedures and repeat work go. The partner writes the playbook, and someone else runs it.

Try this on your next big task

Before your next project or big task, try answering four questions instead of writing steps:

1. What am I trying to accomplish?

2. Who is it for?

3. What does excellent look like?

4. What must not happen?

Give the model that, let it figure out the how, and see what comes back.

The attached cards are the short version you can keep or share.

If you notice a difference, reply and tell me what happened.

And if you want to look at how you're working with AI today, book a call: cal.com/santiago-restrepo/15min.

Santiago