March 2026

If you’re reading this, you’re probably using AI more than you used to — but how has your use actually evolved?

The more I use AI, the more I worry it agrees with me too much … or worse, that I agree with it too quickly.

For me, as AI becomes more powerful, I’m using it in more places more often. It’s becoming a step in almost every process I do.

At the beginning, my use was very simple. I would highlight a sentence and say, “Improve this.” I was often surprised by an LLM’s ability to take a jumble of words and distill something shorter and more meaningful. I’m sure many started to feel like they could put voice to their thoughts.

Then, I was impressed by AI’s ability to turn long articles or collections of sources into tight summaries that made clear why they mattered and what to do next.

Over time, I learned to use AI to help me do things I already did, to the point where it enhanced my ability to do it … or freed me up to do a little bit more.

Now, if I’m doing something repeatedly and not using AI or automation, I assume that’s a problem.

Not everyone feels that way.

For example, this weekly commentary is still primarily written by humans (my son, Zach, and me). As AI becomes a larger part of the production process, Zach becomes increasingly dubious of how I use AI. He worries that AI-In-the-Loop processes impact our writing in ways that we quickly become desensitized to or stop noticing altogether(for example, logical patterns, word choices, common idiosyncrasies, or misplaced confidence).

Too Much of Anything Isn’t Good — Even Agreement

As amazing as AI tools are, it’s well-documented that they can be sycophantic, hallucinate & fabricate, and be surprisingly rigid in their process … if  you don’t have a good enough process in place to manage things like that.

Meanwhile, prompt researchers found that making AI agents ruder resulted in better performance on complex reasoning tasks.

So I wanted an AI that communicates with me like a sharp board member — not a flattering intern. That’s where prompts like the Frenemy Prompt come in.

I saw this on Tech Radar. Here is the basic idea.

Respond with direct, critical analysis. Prioritize clarity over kindness. Do not compliment me or soften the tone of your answer. Identify my logical blindspots and point out the flaws in my assumptions. Fact-check my claims. Refute my conclusions where you can.

The Frenemy Fact-Check

This prompt turns an AI into a hostile‑but‑useful executive reviewer that turns text into decision‑grade clarity by separating what’s said, what’s assumed, what’s missing, and what actually matters.

It’s designed to:

• Take a piece of text (an article, memo, thread, transcript)
• Refuse to trust it
• Separate what is actually said from what is assumed
• Stress‑test it like a skeptical board member, and
• End with a clear decision posture and a best next step

A Fluff-Busting Example:

If you’ve been a reader for a while, then you know I share a lot of links to a wide range of materials. What you see is the filtered list after I’ve read it, tagged it, ranked it, and decided what to share … but what I see is much broader. Some of it is AI slop, conspiracy theory madness, or aspirational thinking masquerading as strategic thinking. I wanted something that helped me sort, sift, and filter what comes across my screen. For example, this week I clicked on this article claiming that new experiments confirm the existence of parallel universes. If I simply asked AI to summarize the article, it would. But this prompt gives me something different. Its response started this way:

Executive Brief:

Posture: Probe — The piece argues that multiple experimental and theoretical threads make the many‑worlds / multiverse idea increasingly plausible; it’s an interpretive synthesis, not a proof.

Biggest risk: Conflating interpretation and empirical demonstration — many claims are inference/speculation built on experimental results.

Next action: Identify the article’s specific factual claims and separate which are quote-backed, which are inference, and which require verification.

That prompt then goes on to identify all the different claims to verify, fluff to bust, and even makes it easy for me to expand the research or reconcile the language. Here is the output of the first step if you are curious.

The point of this article isn’t to share a polished prompt. My production version is long, messy, and customized to my workflow and input sources. However, if you’re interested, here is a basic Frenemy Fact-Check Framework prompt that you can customize.

I’m sharing the idea as a seed — useful on its own, but far more powerful once you make it simple, repeatable, consistent, and scalable.

For context, my current version, 7.0, is over twice as long, has portions that a human won’t understand, and understands me and my needs much better than this seed.

And it was AI that helped me iterate on the prompt until it reached that point.

Creating a Production-Grade Process

The way you do that is by analyzing what you’re doing, both in terms of what the audience sees (front stage) and what is required to reliably produce the front-stage experience (backstage).

Most prompts focus on the front stage and don’t handle the backstage well enough to be reliable in production.

Front Stage vs. Back Stage

Front stage, it looks like: “AI reads something and gives a sharp executive review.”

Backstage, it’s doing something much more important: It’s not focused on “smartness” or “creativity”… it is manufacturing reliability.

Think of it like a restaurant:

• The dining room is what customers see (front stage).
• The kitchen is why the same dish comes out the same way every night (backstage).

A professional-grade Frenemy prompt must include the kitchen spec for decision-grade analysis.

Here are some high-level concepts to consider in a prompt like this.

First Principles of the Prompt

At its heart, the system enforces three laws:

Law 1: Words ≠ Truth

If it’s not quoted, it’s not solid.

Anything not directly supported by text must be labeled:

• Inference (reasonable but not stated)
• Speculation (guessing)

Law 2: Structure Beats Intelligence

There is a difference between could be strengthened by briefly contrasting “clever but inconsistent” vs. “structured and reliable.” My production prompt doesn’t rely on the model being ‘smart.’ It relies on the structure we wrap around it.

It relies on:

• Rigid section definitions
• Mandatory labels
• Forced ordering
• Hard cap limits

This is why it’s long. But, it’s not verbosity — it’s scaffolding.

Law 3: Decisions Are the Point

Every run ends with:

• A posture (Proceed / Pause / Probe / Pivot)
• A biggest risk
• A next action
• A control panel that helps the user choose what happens next

As AI makes analysis easier to generate, it becomes even more important not to automate “analysis for analysis’s sake.” This prompt framework was designed to encourage right actions.

The longer the content and project you give AI, the more likely it is to break protocol and make mistakes. A production-grade prompt like this constrains the AI so it can’t “help” in the wrong way, and blocks hallucinations or fake precision by default. It turns raw text into structured evidence, labels ambiguity clearly, and keeps outputs consistent and stable—even under pressure or long inputs. Most importantly, it keeps humans in control through a clear command interface, which is why it’s far more reliable than the average prompt.

I’d love to hear about ways you’re using AI to improve the quality of your output, enhance your performance, or expand what you believe is possible.