Issue No. 12 - August 25, 2026

When I first started using AI seriously, I read everything I could find about how to do it right. Set clear instructions. Tell it who it is. Tell it who you are. Give it context about your tone, your audience, your goals. The more specific, the better.

So I did. I built it out carefully.

Then I started seeing a different take.

Creators I follow, people who work with AI every day, started posting about pulling back on instructions. The argument was counterintuitive: too many instructions restrict the LLM. You're constraining it before it can actually reason. Strip them back, they said, and the responses get sharper. The thinking gets better.

I didn't know what to do with that. I'd built my whole setup around the first version of the advice.

What made it more complicated is that I've been running a parallel experiment this month. I've been deliberately splitting work across different platforms. Strategy questions to one, creative work to another. Testing which tool actually fits which task. And what I started noticing is that the instruction question and the platform question aren't separate. They're the same inquiry from two different angles.

When I brought the same task to different platforms with the same set of instructions, I got different results. Some platforms responded well to tight guardrails. Others felt more restricted by them. The output quality wasn't just about which platform I chose. It was about how much I was constraining it before it could reason. That's when the instructions question stopped feeling abstract and started feeling like something I actually needed to figure out.

What I've noticed so far is that there seem to be two different kinds of tasks, and they don't respond the same way. Some tasks call for consistency. A specific voice, a particular format, a defined output. Those benefit from guardrails. The instructions keep the output in bounds I already know I want. For that kind of work, clear instructions genuinely help.

But tasks where I need actual thinking are different. Strategy. Problem-solving. Working through something I don't understand yet. When I bring heavy instructions into that kind of work, the responses feel managed. Like the AI is trying to satisfy the parameters rather than actually reason through the problem. Pulling back gives it more room. The outputs reflect that.

I've also started thinking about skills differently in this context. A skill applied to a response isn't the same thing as an instruction layered over the whole system. That distinction matters more than I initially gave it credit for.

What I don't have yet is a clean framework for which tasks fall into which bucket. And the more I sit with this, the more I think one may not exist. Not a permanent one, anyway. The answer is going to look different depending on your specific work, your specific objectives, and how the tools keep changing. What works for my setup at JB Sales isn't necessarily what works for yours.

The creators I follow are still figuring it out too. Which means the most important thing isn't finding the right answer. It's staying in motion. Find the people, newsletters, and podcasts you actually trust on this. Not the ones selling a definitive framework. The ones who are honest about what they still don't know. Those are the ones worth following through something that keeps moving.

That's where I am with it. Watching, adjusting, and paying attention to who's worth listening to along the way.

Where this lands in the three buckets

Immediately actionable: Look at your current AI instructions and ask honestly: are they producing consistent, useful outputs? Or are they making the responses feel boxed in and managed? If it's the second, try pulling back on a task where you're not worried about consistency and see what changes. That reaction is information.

Start thinking about: Whether the work you bring to AI tends to be consistency-driven (you need the same kind of output reliably) or thinking-driven (you need the AI to actually reason through something with you). The setup that works well for one can actively hurt the other. The platform you choose and how much you constrain it are both part of that equation.

Next phase: A working rule, not a framework. Just a distinction you can apply before you start. Which bucket does this task fall into? What does that mean for how much I constrain it, and which platform I take it to? Build in the expectation that it will keep changing. And find the voices you trust to help you keep up - people, newsletters, podcasts that are honest about what they still don't know. The goal isn't to arrive at a final answer. It's to stay intentional as the answer keeps evolving.

Meghan Brenner is COO at JB Sales and founder of The Operator's Notebook. The Muddy Middle publishes every Tuesday.