Caveat Promptor

Going back out on my own means doing a lot of things I've never been particularly good at.

I don't know much about accounting or law. Sales has been part of nearly every job I've had, but it was never really my job. Then I started running a practice, which meant I needed to do it all anyway.

So AI has been my co-worker. I've used it to set up invoicing, draft proposals and walk me through contracts before I sign them. I know it isn’t legal advice, and it won’t replace an accountant. But it gives me enough general knowledge to get started without guessing.

For many people, this may be the most valuable thing about generative AI. The headlines tend to focus on models writing code and passing professional exams. But what's changed my week is something more ordinary. 

Until recently, getting general knowledge often meant paying someone to explain it or going without. Now I can get it at eleven at night and keep asking follow-up questions until it makes sense.

But there is a catch. AI sounds just as confident when it's wrong as when it's right. And I'm leaning on it most heavily in areas such as accounting and law, where I know the least, and am therefore least equipped to spot an error.

There's an old Latin phrase for this: caveat emptor, which translates to let the buyer beware. Nobody is selling me these answers, though. I'm the one asking the questions. So perhaps the most appropriate version for this is caveat promptor: let the prompter beware.

The model isn't going to sign the contract. I am. If it gets something wrong, I'm the one who has to live with the consequences. 

I mostly use the AI to get oriented and learn enough vocabulary to ask a professional a better question. When the stakes are high, my job is to know when to call an actual accountant or lawyer.

There is also a distinction between what the model knows and what I know. It knows how proposals are usually put together. But it doesn't know what a particular client is worried about. I do, because I've been in a lot of rooms with people like them. Put the two together, and I can produce something I would have struggled to get started on a few years ago. It keeps me from staring at a blank piece of paper.

AI knows what everybody knows. What you do with it is on you.

The Alan Bot

In April 2025, in GPT Doesn't Know What I Know, I wrote about the difference between explicit knowledge, which can be written down and fed into a model, and tacit knowledge, which lives in people. 

I also imagined someone going through my old emails and notes after I left a job, feeding all of it into an LLM, and calling the result "Alan Bot." I said that if anyone was building something like that, I might write them a check.

Nobody took me up on it. Eventually, I started building one myself.

In They Skipped The Local Streets, I described the project I use to write this newsletter. It contains every issue I've published, every blog post, and every LinkedIn post going back 15years. I thought of it less as a writing tool than a memory I could query. 

I didn't have a word for it then. The industry now calls it a company brain.

Companies are increasingly building this system. I've been using Glean at work. It brings together information scattered across company documents and conversations, allowing me to get answers based on what the organization actually knows rather than relying on the open internet. That makes sense for a company.

Personal AI is moving in a different direction. Instinct, an invite-only assistant you can text or call, raised $1 billion at a $10 billion valuation in September. Meta launched Muse around the same time. You tell these assistants what you need, and they can book a trip or cancel a subscription you’ve forgotten about. 

These are assistants for getting things done. But I haven't found a personal version of Glean: something that pulls together what I know, including how I assess a client and which of my gut calls have held up over the years, and makes that knowledge available to whichever AI model I'm working with.

Right now, every AI tool I use has its own version of me. None of them talk to each other. I'd like one place where the whole picture lives, with me deciding what each tool gets to see.

I've put the most into the rules of my newsletter project. No reference to an old issue goes into a new one until I've found the actual sentence in the actual file. That is because the model has invented callbacks before. There is also a list of phrases I don't use. Most of those rules exist because something went wrong once.

In a way, that's caveat promptor written down. It's a record of where I've learned not to trust the machine. 

Companies can formalise these lessons and put them in a policy. Most of us, on the other hand, just carry it around in our heads.

Berkson's Bits

The difference between being persistent and being annoying is the homework you do to be relevant and the respect you show in the process. 

What I'm Listening To...

Here’s a collaboration I didn’t know I needed. Herbie Hancock and Stevie Wonder on St. Louis Blues. 

I don't know yet whether this becomes something people buy or something each of us builds. Mine is still pretty rough. 

But building it has reminded me that so much of what I know remains unwritten and undocumented. It's also one of the reasons I started this newsletter.

If you've built something like this for yourself, what did you put in it first?

Looking forward to continuing the conversation...

Alan