Anthropic recently told developers to pull back on heavy system-prompt instructions. The models have gotten good enough that in some cases, removing instructions actually improved results. Load a model up with too much setup, and you might be working against yourself.

I build products in this space, so the implication is direct.

What a system prompt actually is

When you set up an AI tool, most of them have a behind-the-scenes layer where the product creator has written a set of instructions. Think of it like a briefing document the AI reads before you ever type a word. It shapes how it talks, what it focuses on, what it knows about your business. The more detailed that briefing, the more "customized" the experience feels.

A lot of AI products, including Business Brain, are built on this idea. You put your business information in, the tool builds you a structured briefing file, and every time you use the AI after that, it already knows who you are and what you're doing. That's genuinely useful. The AI stops being generic. It starts sounding like it works for you.

Anthropic's guidance doesn't say that's wrong. It says the assumption that more instruction equals better results deserves a second look. The models are stronger now. Some of what used to require explicit instruction, the model handles on its own.

What that means for something you bought

If you paid for an AI setup product, you're probably asking a fair question right now. Did the value you paid for just get patched away?

For most people, the answer is no, but the reasoning matters.

The part of a product like Business Brain that holds up over time is the business information you put into it. The AI knowing your services, your customers, your tone, your pricing, your differentiators — that's durable. The model keeps getting better, which means a well-built context file about your business actually gets more useful over time, because a smarter model can do more with the same information.

The part that's more fragile is any layer that was compensating for what the model couldn't do on its own. If a product spent a lot of its instruction budget telling the AI to "be concise" or "always use bullet points" or "respond as a helpful assistant," that layer matters less now. The model already does most of that.

Good AI setup products were always about capturing your business knowledge. If the product you bought did that well, you're fine. If it mostly built scaffolding to compensate for a weaker model, the update hurts more.

I rebuilt Business Brain's core UX around visibility after realizing this

I've made several UX changes since the early version shipped. The one I keep coming back to is visibility. In the original version, you'd go through the whole setup, answer every question, and then get a bundle of files. But you couldn't see what the system actually concluded about your business before downloading. You couldn't tell if it had pulled the wrong thing, made a bad assumption, or missed something important.

That bothered me. You'd handed over a bunch of information and had to trust the output without inspecting it.

The current version shows you the conclusions before you download anything. You can read what the system thinks your business is about, disagree with any of it, and adjust before the file gets built. You're signing off on the content rather than hoping it got it right — and that changes the relationship between you and the tool.

There's also a voice option now, so you can talk through the setup instead of typing. And the download is one clean file with all the individual pieces inside it, instead of a mess of embedded files you couldn't inspect.

These changes don't have anything to do with Anthropic's guidance on system prompts. They came from someone using the product and pointing out the gaps. But they point in the same direction: what you should be able to see and verify is your business knowledge, and the tool should get out of the way of that.

The last round of user feedback changed how the product works

Anthropic's calibration is this: don't assume complexity is the same as quality. Don't pile on instructions because it feels more rigorous. Test whether the model needs them.

That's a good principle for evaluating any AI tool you've bought, too. Ask what's actually in the system prompt. Is it information about your business, or is it scaffolding to compensate for a weaker model?

If it's mostly the first, the product will age well. Models improving means your business context does more. If it's mostly the second, each model update quietly erodes what you paid for.

For Business Brain, I'm watching this closely and building toward more durability, not less. The context your business generates is yours. It should travel with you regardless of which model you're talking to or what Anthropic ships next quarter. That's the goal I'm building toward.

If you want to see what the current version looks like, it's at daringstrategy.com/business-brain. And if you've already been through it and have notes on how it felt, send them. The last round of feedback changed how the product works. That's what happened.