AI business writing coach Nate B Jones explains why most AI-generated documents fail and how to fix them
A solo presentation by Nate B Jones on the principles behind effective AI-assisted business writing, including a walkthrough of a high-quality meeting notes prompt.
Summary
Nate B Jones draws on 200 hours of experience teaching AI writing to argue that the core problem with AI-generated business documents is not the AI model itself, but organizations' inability to articulate what good writing looks like. He identifies a "specification bottleneck" — the shift from asking "how fast can I write this?" to "how clearly can I articulate what I need?" — and argues that AI amplifies ambiguity rather than reducing it. He also makes the case that evaluation must be scaled using AI, not just generation, and that organizations need to define failure modes explicitly, not just success criteria. The video closes with a walkthrough of a structured meeting notes prompt that demonstrates these principles in practice, contrasting it with the generic, low-intent output produced by most AI note-taking tools.
Key Takeaways
FULL TRANSCRIPT
The core problem with AI business writing
Nate B Jones: AI has dropped the cost of business writing to nearly zero. And most of the businesses I work with or know about are drowning in AI documents and have huge issues with AI business writing. This video is how you troubleshoot those — what the principles are for using AI well in business writing situations. And then I'm going to walk through an actual example of a prompt I'm using that sets a much higher bar for AI business writing than I've typically seen.
So first, let's get into the principles. The first thing I want you to keep in mind overall is that the real bottleneck in AI-assisted writing is never the capability of AI. People think it's the model. It's not the model. Don't let anyone tell you it's the model. It's organizational ability to articulate what constitutes good work. And typically that means we've relied on individuals to have instincts for what constitutes good work instead of actual structured information about what constitutes good work.
AI forces tacit knowledge into explicit standards, and that is very, very hard for most businesses. You cannot rely on "I know it when I see it," because AI cannot read your mind. It is not that good. It's never going to be that good. Every quality criterion needs to be concrete enough to specify, to test, and to verify. That is the only way through on the business writing side.
And if people say, "Well, I don't have time for that" — I've got to ask you, do you have time for the business writing you're drowning in? Because I have lost track of the number of people who say they cannot keep up with the business writing, that it is too much, that people are sending them AI slop at work. Well, the quality criteria need to be defined to make that go away.
The specification bottleneck
So what does that actually look like? First, understand that there is a specification bottleneck. The barrier is not "how fast can I write this?" anymore. It is "how clearly can I articulate what I need?" Every time you have ambiguity in your specs for a document, that is amplified through generation. It is not reduced. People sometimes think AI can reduce ambiguity by adding detail, but anyone who's worked with AI a lot will tell you it doesn't reduce ambiguity — it enhances it. And when it adds helpful-seeming detail, it makes things worse.
So the organizations that succeed are actually not those with the best writers. I know that's very counterintuitive. They are those who can articulate quality standards explicitly enough to encode them in prompts that you can work with. Now, writers can absolutely help with that. Good writers are hard to find. But we are moving from raw ability to generate text toward building a congruent prompt framework or a congruent template as a goal. We're getting into a world where we just need to specify our requirements really, really clearly. It reminds me of wearing a product hat and defining product requirements — very similar, except now your product is the document.
Scaling evaluation with AI
You also have a fundamental evaluation issue because of the volume of prompts being run. I've talked in the past about how we have this issue with résumés — really, it's with all knowledge work in the business. We don't have time to evaluate everything that gets done, which means we have to figure out how to scale evaluation. That is one of the fundamental challenges for businesses that want to go faster.
And I believe firmly that scaling evaluation means putting AI on the evaluation side, not just the writing side. I want to talk about that a little more because I think we miss it. It is absolutely possible — I've done it. I'm writing a prompt for this article to talk through how you evaluate and build a Claude skill that helps you evaluate. You can make your job so much easier if you are just willing to let AI take a first pass and give it really clear requirements on what good looks like.
Information architecture problems
I also want to go beyond just "you need to specify." The other core issue is that we typically have longstanding information architecture problems in our documents at work that we just paper over — and that we are not going to be able to paper over when AI is writing. Because fundamentally what AI does is expose information asymmetries, informational vagueness that previously hid in a lot of our writing, because people wrote it and you just assumed they were doing their best thinking.
One of the good things about the AI age is we now don't assume that everyone's doing their best thinking, which means we critique more — which is actually healthy.
So when we're critiquing, I want us to think of a few things in this information architecture bucket. One: documents really ought to be written for goals and decisions, and they aren't always. If you can't tell whether this document enables person X to make choice Y, if you don't know what it looks like and why you're reading it, it was going to be useless anyway — but now we blame the AI. Your goal at this point is actually to get rid of the vagueness you tolerated in the past and define the informational architecture of the document.
Your structure is the business logic, not just a template. So many times, if someone gives you a product requirements document, a business memo, or a press release, when they give it to a human, you get a template. You actually don't get the logical underpinnings of the document, and people end up learning those from experience. Templates just let you fill in the boxes without thinking. And when you hand AI only a template in the prompt, that is what you get. That is why so often when I'm called in to help with business writing prompts, people say, "I don't know what I did wrong. I gave it the template and it filled out the template and it's terrible." Well, you didn't give it the business logic. You didn't give it a decision interface to work against. You didn't give it a goal for the document. That's why the writing is terrible. If you don't give it that intent, the business writing is going to fail.
Defining failure modes
So the other piece here — and this is very counterintuitive — you can't just give it a goal and give it business logic. In my experience, you also have to give good failure tests. You have to insist that you know what bad looks like. Isn't that counterintuitive? But if you're trying to tell the AI how to do something well, it really helps if you have five to seven examples of the kinds of quality problems you have with these kinds of documents.
For example: "This technical specification document is really over-specified on the design and insists on a microservices architecture when we don't use that." Great — that's a failure example. Or: "This press release is way too hyped and doesn't respect the actual product capabilities." Or: "This executive summary is too vague — I need more specificity." Understand where your organization today fails to communicate information, and you will understand how to work with AI to write better.
This is — and I'm going to repeat it — a people problem at root. It is not the model's fault. It is our ability in organizations to communicate intent clearly that is governing our ability to work with AI, and we're not doing it well.
Voice convergence and information loss
I want to tease out some of the organizational dynamics too. Specifically, one of the things I've noticed that's subtle but painful: we are converging on a single voice because of AI, and that is leading to informational loss in business systems. We have an AI default voice, and too few people understand how to push that voice into something that communicates their intent clearly. I'm not talking about style here. I'm talking about the ability to communicate clearly what really matters.
The default AI voice is diplomatically hedged. It's pseudo-comprehensive. It's stylistically extremely bland. And you don't have the ability to carry conviction with that voice. If you want to make a bet, you don't have the ability to articulate real specificity — but in the same document, to also say "this area is vague and uncertain, and I want to admit that upfront." Good quality writing has that range. And AI, if you just prompt it vanilla, does not. That leads to critical information loss. And that is part of why businesses feel like they're drowning — the information is not high quality.
And I'm going to say again: it is absolutely possible to do better. You can make high-quality documents with AI.
Iteration diagnosis
The last thing I want to call out before I show what I mean is iteration diagnosis. Very simply, we need to diagnose the failure of people to iterate well with writing. People are trying to say "make it better" on their business documents and that is all they're writing, and it's not working. But no one knows how to do it better unless they're educated. What they don't realize is that it is a people problem to communicate intent, and that they have to specify their intent more clearly if they're not getting a draft they like.
So I'm going to come back to the core of this issue and then show you how I'm addressing it with a specific prompt. The thing to remember is: because AI-assisted writing is exploding and organizations are drowning, we have that AI generation problem. The cost of information is zero. We need to therefore put a premium on our intent. Otherwise, we degrade informational signal through our businesses, it becomes hard to make decisions, we feel like we're drowning, and it has real career implications and real dollars-and-cents implications.
I am passionate about good business writing. I love it. It is getting hard to find because people don't know how to prompt. Let's get to an actual example.
Walkthrough: a high-quality meeting notes prompt
This is an example of a prompt that I think is high quality. It is also designed to be modular and changeable so you can make it the way you want. This is for meeting notes — the simplest possible case. I have a bunch of other prompts for more complex documents.
Meeting notes are overlooked because most of the time if you go into a generic AI transcript and get meeting notes, you just get a generic summary — very vanilla — of what was there. I wanted to be more opinionated because I wanted to carry through the principle that you need to have intent around what you're doing.
So the prompt has context, date, attendees — which should be pullable from the meeting note raw input — purpose of meeting, input provided where you can paste your transcript, and then you're asking for a very specific output. Your goal here is to create notes that help the team execute on what was discussed. There is a specific goal for this. The notes are used in a specific context — you can decide where they're used.
You then have a required structure: Did you make decisions? Do you have action items? Are there open questions that were discussed? What were the key discussion points? The vanilla notes I get — and I love Granola, I love Otter — those AI note tools do not do this. They do not help you encode intent. It is up to you to bring this level of clarity. The AI won't bring it.
You have constraints: a total length you have to keep to, decisions that must define an owner by name, action items that cannot be vague, open questions that must have assigned owners. You may not include pleasantries or general discussions. You may not infer. You may not guess. And then there are validation quality checks: every decision must have a named decision-maker, every action item must have an owner, no action item is allowed to be vague, open questions must have assigned owners. If any check fails, revise before outputting.
Is this perfect? No prompt is perfect. Is this going to get you a long way on intent? Yes.
And then I get into why the prompt works: it communicates purpose, it communicates structure as logic — and you can change that structure if you want a different intent. There's an evaluation mode. There's a failure mode. And I include guidance on how to customize it for your workflow — you can change it to your organization's voice, you can have different meeting types, you can have a sprint goal instead, you can have different failure modes.
And then I give an example of what a good output looks like versus a terrible output. The terrible output looks very similar to what I get generically from most AI note tools. Frankly, ChatGPT launched a meeting notes feature that looks a lot like that top example. This is part of how I know we're losing good quality business intelligence. The better example is going to be much more informative for the business than the generic one.
Closing argument: humans must define what good looks like
Please — I have a bunch of these prompts. I don't care if you use my prompts or not, but please put intent into your business AI writing. That is the key. If prompts help you scale that across the business, if Claude skills help you scale that across the business, great. But there is no substitute. You cannot get away from the need for humans to define what good looks like for AI and to define requirements.
And to be honest with you, that is the thing I'm excited about. We have sat for a long time with the assumption that human best effort is kind of the bar for documents. At Amazon, we had a bar for documents that floated around based on the best human writer on the team or in the department. You don't have to have that anymore. You can have a really consistent, high-quality bar, and you can know whether someone is writing to that bar or not.
People ask me all the time: "Does this mean people won't think anymore?" I dare you. Are you going to think less if you go through this process? If you actually define intent for your business with writing, no — you are going to think more. You are going to think harder. You're going to have to work harder to communicate all of this to people because so much of it was vague and lived in people's heads. Well, not anymore.
And the reason you're going to have to do this is because the alternative is not what we had pre-2020, where everyone wrote everything. The alternative is AI slop forever — because AI is out of the box, everyone's using it, and everybody I know at work who is drowning in AI documents, which is a lot of people, that is not going to stop. The people making them aren't going to stop because they think it's productive. We need AI education that emphasizes quality and emphasizes the different ways we need to think.
I hope this video has helped you think about how our brains need to change to communicate effectively with AI when we are writing.