This afternoon I was talking with Perplexity Teammates, and it made a good point. Over the past three years, AI tools have gone through many cycles. One thing I’ve been saying all along has stayed the same.
Put your own knowledge and experience into AI so it can perform beyond a general-purpose baseline, and keep ownership of what you build up along the way.
Knowledge bases, Skills, and workflows are recurring themes across the hundreds of videos and articles I’ve published. A knowledge base gives AI domain knowledge. Skills and workflows give it domain experience. Both help AI perform beyond the general-purpose baseline.
Why should AI perform beyond the general-purpose baseline?
It comes from a basic judgment of mine. AI will become widespread, and it will become the average level of capability across society. When everyone uses the same models and the same agents, knowing how to use AI will cease to be an advantage, much as knowing how to use a search engine is ordinary today.
Models know what’s public and general. They have no way of knowing the information in your industry that never made it into public materials, exactly how you do a piece of work, or what people in your field consider a sound result or an embarrassing one.
So if you give the same model to a seasoned practitioner and a beginner, and they both ask it questions directly, they’ll get roughly the same 60 out of 100.
Personal agents will make this even more apparent. Platforms will get general-purpose memory right, covering who you are, your schedule, and your preferences. Those will soon be standard features everyone has. Your professional expertise is the part a platform can’t supply.
The real difference, then, comes down to who can get AI above 60 in their own field. The people best placed to do that already have expertise in their fields. That includes the folks in our newtype community.
How can AI perform beyond the general-purpose baseline?
Many of you have already begun doing exactly that by building an AI professional layer.
An AI professional layer organizes the knowledge, experience, and judgment you’ve accumulated in a field into materials, methods, and standards that AI can use for a specific task.
An AI workbench is one way to house an AI professional layer. An agent can draw on that layer to carry out tasks. Workbenches and agents will come and go. The professional layer is what you keep building.
The AI professional layer has three characteristics.
First, you own it. You organize and save the materials, methods, and standards yourself, and can keep them across platforms. Switching tools takes some adaptation and testing, but you get to keep the work you’ve already done.
Second, it needs maintenance. The results of real tasks and the corrections you make will reveal missing materials, unsuitable methods, and unclear standards. It gets better through testing and maintenance.
Third, it serves real work. Its purpose is to help AI complete specific tasks. A collection of notes or prompts alone can’t serve that purpose.
What does an AI professional layer look like?
The AI professional layer consists of three aspects of expertise you build up over time. They are knowledge, experience, and judgment.
Knowledge — what AI knows
Knowledge covers what the model lacks or gets wrong about your field.
That could include industry standards, internal documents, project records, and client background. It could also include notes you’ve accumulated over the years, research reports you’ve read, and cases you’ve organized.
Knowledge bases, wikis, and well-organized reference materials are ways to hold that knowledge.
The important thing here is how well you organize it. Use clear formats, make the sources traceable, and flag outdated information promptly. For AI, a pile of disorganized PDFs might as well be no materials at all.
Experience — how AI works
Experience is how you do a piece of work. What comes first? What comes next? Where do you need to pay special attention? Which common practices are traps? So much of this is in your head, and it can feel obvious to you. The model has never learned it from you.
Skills and workflows are ways to capture that experience. Spell out the steps, the points to watch, and the examples, then package them so AI can call on them. It can then do the work your way instead of falling back on the average approach it finds online.
Judgment — how AI sets goals, makes trade-offs, and checks results
Judgment is the set of standards and trade-offs you apply to a piece of work. What does this task actually need to solve? Which of several approaches should you choose? What counts as correct or good? When does the work need to be redone? When should it stop and be handed to a person? Judgment runs all the way through, from setting the goal and choosing an approach to checking the final result.
The idea is for AI to make trade-offs and check work against your standards, and to know the limits of what it can decide. You retain the final call.
You can capture judgment in task briefs, checklists, acceptance criteria, and examples of what passes and what fails. The biggest difference between an expert and a beginner often comes down to knowing what matters and spotting what’s wrong at a glance. Write those things down, and AI’s output stands a better chance of going from “looks pretty good” to “ready to deliver.”
Written judgment can only cover situations you’ve thought through. There’s no way to fully reproduce the judgment an expert exercises on the spot in a new situation. That’s when a person should take over.
How the three fit together
Think of it this way. Knowledge is the raw material, experience is the method, and judgment is the standard.
Knowledge alone can make AI’s answers more accurate, but that may still leave it unable to do the work. With experience, it can do the work your way, but it may get it wrong or do work that shouldn’t be done in the first place. Add judgment, and it knows where to put its effort and what it takes to call the job done. Many people find AI “astonishing sometimes and wildly off base at others.” A common reason is that they haven’t spelled out how to judge the work.
I can see the direction now
That’s where my conversation with Perplexity Teammates led.
The people in our newtype community and I are past introductory topics like how to operate AI tools. We’re interested in the more advanced question of how to pass everything we’ve accumulated over the years to AI and help it perform above that 60-point baseline.
The AI professional layer is what we’ve been working out all along. Its three pillars are knowledge, experience, and judgment. They also happen to be among the AI industry’s most important advances in recent years.
Now that I have that direction, I know how to organize my content and products going forward. Those 200 dollars for Perplexity Max were well spent!