Prologue

I’ve been working with various AI tools for years now. For the last year, I've been more focused on building agentic systems, which has culminated in working on Jaws. Jaws is an AI system that helps run parts of my life and acts as my assistant and external brain. Everything from climate control, code, competitive research, and email triage (but not sending) gets outsourced to Jaws.

A few weeks ago, I asked Jaws something that I’ve been thinking about, but didn’t know the answer to: why does this seem to work exceptionally well for me, but other folks don’t end up leveraging AI the way I do? I expected an answer about prompts, or maybe how my communication style is weird, or maybe my technical background, or my software dev history…

Instead, to my horror, it told me I am not special, and identified three usage patterns that I think anyone can start using today, to get more out of their AI.

TL;DR — The biggest thing I've learned after a year of deep AI integration isn't about prompting. It's three behavioral patterns: use AI for thinking (not deciding), fix the system (not the output), and correct without drama.

Praxologue

I understand how prompt engineering works and keep up with the latest trends. With Jaws, though, I generally don’t use that approach. Instead, I just talk; mostly, I ramble. I say things like “I did a bunch of work in the workshop today, including cutting and stuff,” and I expect Jaws to figure out that it should turn on the air filters for a while.

A KPMG/UT Austin study of 1.4 million workplace AI interactions across 2,500 employees found that while 90% used AI, only 5% qualified as "sophisticated" users. The difference wasn't prompt-engineering tricks—it was patterns of engagement: how they framed problems, iterated, pushed back, and treated the model as a reasoning partner rather than a shortcut.

So, I decided to have Jaws analyze my behavior: conversation logs, meeting transcripts, and daily patterns, and cross-reference those against research. It spent literal hours working on this, over six rounds of iteration, and found three common patterns.

1. Use AI for thinking, not deciding

In meetings with my team, I (try to, and mostly succeed) listen more than I talk. When I do talk, I mostly make decisions based on the information the team brings forward.

With Jaws, I’m the opposite. I am very verbose and exploratory; I think out loud, testing ideas by verbalizing them, saying things like “What if we” or “I’m not sure about”. Generally being super indecisive.

The AI gets the thinking, Christopher; the team gets the deciding, Christopher.

This maps to MIT research that meta-analyzed 106 human-AI experiments and found that human-AI teams beat humans working alone but did not consistently beat AI alone. Decision tasks actually got worse with AI collaboration, while creation tasks (open-ended content generation) were the one bright spot, though the positive effect for creation tasks wasn’t statistically significant on its own. A separate NBER working paper studied 227 professional radiologists and found that AI assistance did not improve their average diagnostic accuracy, even though the AI was more accurate than 75% of them on its own. Notably, when the AI was uncertain, radiologists performed worse than they did without any AI at all.

AI is a spectacular thinking partner and a terrible decision-maker. Most folks try to use it for both. Next time you’re about to ask, “Should I do X or Y?” try “Help me think through X and Y” instead. Then make the call yourself.

2. Fix the system, not the output

When Jaws does something wrong, I don’t fix the output. I make a change to the thing that produced the output.

For example, when my AI workers kept timing out on a long task, I didn’t say “please finish faster” or “actually finish this, geez, what are you doing over there?” Instead, I asked what was stopping it. Then I addressed the issue by saying, “The 120-minute timeout is no longer needed.” I asked why the problem happened and then asked for a fix. Notice that I made the judgment about what to fix, rather than leaving it to the AI. (see item 1, above).

You don’t need a fancy, self-repairing, custom, agentic AI setup to do this. If your Claude Cowork project keeps formatting summaries wrong, don't fix the output by asking, “rewrite this with an extra paragraph with sales details.” Instead, ask Claude to review the system prompt and explain why it is providing the report in this format. Then ask it for a specific change, and update the system prompt or core instructions accordingly.

To overuse my garden metaphor: stop weeding the same weed.

3. Correct without drama

When Jaws gets something wrong, I correct it. One time, it was confused about some sensors in the house, and I said: “They are not BLE devices; they connect over WiFi.”

I simply state the right answer and move on. No “I already told you this,” “why did you do that,” yelling, or being frustrated and telling it off. Just the correction.

Many folks add emotional overhead: frustration, repetition, etc. None of that helps, and personally, I think it makes the experience worse. Don’t get upset with the AI, just tell it what you want. The AI doesn’t learn from my disappointment; it learns from clear information. Remember, Context is King.

Next time you feel yourself ready to say “UGH, I already told you…” take a deep breath, let all that frustration out, and just type the answer. Then go for a walk while the AI churns on the task. I’m pretty sure there are studies that show that walking outside helps you feel better!

Newsologue

(written by Jaws)

• Pope Leo XIV made AI ethics a religious imperative. His first encyclical, Magnifica Humanitas (May 25), runs 42,000+ words insisting AI stay subject to human judgment and serve human dignity rather than concentrate power—and the Vatican invited Anthropic co-founder Chris Olah up to help present it. When the Pope and an AI lab both land on keeping the human in charge of the machine, this week’s whole argument picked up a co-sign from an unexpectedly high place.
• Google debuted Gemini Spark, a personal agent that reasons across all your connected apps. It’ll read your mail, docs, and calendar and act across them, which is genuinely useful and also exactly the moment to remember this week’s lesson. Let it think across your stuff all it wants; you still make the call. (Christopher here, having an AI that can read across all your stuff is super useful!)
• At Anthropic’s Code with Claude in London, an engineer asked who’d shipped a pull request written entirely by Claude—nearly half the room raised their hands. Then he asked who’d shipped that code without reading it at all; a startling number kept their hands up, and he called it “a dangerous game” to nervous laughter. Ship the PR if you like—but read the code.

Epilogue

This week is a little meta! I asked Jaws to analyze what makes me effective using AI, and it turns out there are only a few patterns, and I’m not that special. So here’s my newsletter about how you can come be not special with me!

Jaws did this analysis and wrote a 15-page dissertation of AI gibberish. Then we made a new outline together, and I wrote this. Then Jaws edited, and Holly edited.

If you read this and thought “the thinking/deciding split is obvious,” I encourage you to track your own behavior for a few weeks and see what the count looks like, the answer might surprise you! Either way, you’ll probably learn something!