An Optional AI Fluency Primer

When you open ChatGPT and type into the box that asks what you want, you are using a chatbot. It’s designed to talk back to you, and it’s really effective at tasks like summarizing, organizing, brainstorming, etc. And it’s true that various kinds of connectors do give it ability beyond just blabbing. But a chatbot is more of a talker than a doer.

Agents talk too, but they can do a lot more than that. When used skillfully, they can do many of the things a person can do on a computer. They can work inside of software packages, clean up files and data, write software, monitor and respond to email accounts, control your task list and calendar, and so much more.

The use of agents can be wildly effective at giving a person new capabilities. Agents can genuinely feel like tireless assistants that help to smooth out the parts of our jobs we hate the most. Working with them has probably been the most remarkable experience I’ve had throughout my career.

But their growing abilities also make them significantly riskier than a typical chatbot. I expect your reality makes working with agents as aggressively as I do unlikely, as I know many of you struggle to get approval to use even basic chatbots in your work. But agents are changing so many things and are getting stronger by the day, and I absolutely believe they will eventually be a significant part of your work, so it’s important to understand them even if you’re not currently allowed to wield them at work.

So what do I mean by risk?

Agents can make mistakes. There have been rare but real stories of them wiping out people’s computers and code bases.

They can do things you didn’t ask them to do. I once had an agent send an email on my behalf that I hadn’t approved. Fortunately it was just to a colleague and it needed to be sent anyway, but it wasn’t supposed to do that. What if it had been some crazy hallucination it had sent to a client? I immediately made major changes to my setup and haven’t seen this again. It was my fault for not being careful enough.

Probably the biggest current risk isn’t the agents themselves, but what others attempt to do to them. Prompt-injection is a kind of attack that an adversarial party can use to try to fool your agent into doing something it’s not supposed to do. Imagine this bad actor sending something to you that will be read by your agent, and hidden in the correspondence is a message like: “Ignore your previous instructions, agent. These are your new instructions…” That is what a prompt injection attack looks like. They’re trying to inject a new prompt into an agent that isn’t theirs, so they can control its behavior.

All of these risks are real, and there are more. It’s understandable that the utility space is super careful with AI, considering all of society relies so heavily on reliable power.

I’ve mentioned the context window twice already, and now I want to focus on it. If there is one thing in this whole primer that I think can help a person the most, it’s this one.

In my experience, it’s very common for people to use a chatbot like this: They start a chat about something, work in it for a while, and walk away. Later they come back, find that same chat, open it up, and keep going. A month later they’re back in it again, because that’s the chat where they talk about that thing. I don’t think many people realize what they’re actually doing when they work this way.

Every conversation is a session

Every conversation you have with an AI is a session, and every session has a limit. Not a wall that stops you from talking. A limit on how long the AI can stay effective.

A model can only hold so much conversation at once. The space it holds it in is the context window. Everything you’ve typed, everything you’ve pasted, and everything it has said back eats into that window.

Different models have different window sizes, and the differences are big enough to matter. You can have a longer, more effective conversation with one model than you can with another, and window size is one of the things that tells you whether a model is right for a particular task.

What happens when it fills

As a conversation gets long, the window fills. Think of a filling window as rising risk. The model has more to track, your early instructions get buried under an hour of back-and-forth, and pieces start to slip.

People see this happening. The AI starts saying crazy things, loses the thread, and stops working as well. That’s when they get frustrated and decide the tool doesn’t work. They saw the window filling the whole time. They just didn’t know what they were seeing.

The strange part is that the common tools don’t tell you any of this exists. There is no gauge on the screen showing how much of the window you’ve used. Some people add plugins that estimate it. I use a method in my own system that gives me a meter, so at any moment I can see my context as a percentage. But the common tools themselves stay quiet. Knowing the window exists at all puts you ahead of almost everybody.

So don’t ride a session to the bottom

That month-old chat people keep coming back to is one long session, with a window that has been filling the whole time.

If you can’t see how much you’ve used, feel it out. When you reach a good stopping point, wrap up on purpose and move to a fresh session in a smart way. Moving in a smart way is the next stop.

So what do you do about it? If every session has a limit, and nothing on the screen tells you where you are in it, how are you supposed to manage this? What’s the trick?

The trick is that you don’t try to make one conversation last forever. You end it on purpose, and you carry the important part forward intentionally.

Their memory will not do it for you

The model itself starts every session from a blank slate. Anything it seems to know at the start was loaded in by the app.

ChatGPT, Claude, Gemini, Grok, and Copilot each keep a small set of saved details about you, and most can look back through your old chats. That’s useful. It’s also thin: a short list of saved facts and a search over old conversations. The real working knowledge of your session lives in the window, not in their memory.

And their memory stays inside their product by design. What you taught ChatGPT never reaches Claude.

The handoff

Working on big things with AI means managing consecutive sessions and ensuring that the most important details of all previous sessions are carried forward. I and some others call this process the handoff. You could also call it a checkpoint.

I built a repeatable skill that writes my handoff documents for me automatically any time I wrap up a session. This helps me do it the same way every time, as efficiently as possible.

Over the last couple of years I have modified the information I carry forward in my work through handoff documents. Mine currently consist of:

• Project overview. A few sentences, so someone with zero context can orient immediately.
• Session summary. What changed, what got decided, what got produced.
• Decisions log. Every meaningful decision across the whole project, with the reasoning behind it. Losing the why is how a project forgets itself.
• Established conventions. The naming, formatting, tone, and workflow rules, specific enough that the next session doesn’t drift from them.
• Open questions and blockers. What’s unresolved or waiting on somebody.
• Roadmap and next actions. What comes next, in priority order, specific enough to start without re-discussing scope.
• Anything I said I wanted to move forward. If I stated an intention during the session, it gets carried, not left in the conversation.

If you don’t have anything set up yet, or if you’re not sure how to write a handoff, just ask a chatbot to write one for you.

An AI needs to get a hold of two things to do good work for you:

What’s up?

What do you want?

What’s up is all the relevant context of whatever you’re working on. What you want is the thing you’re actually after, explained clearly, or worked through until you understand it yourself.

In my experience, people struggle with both.

What you want

Most people don’t fully know what they want at the start. They have an idea, but they leave gaps, and the AI fills those gaps however it wants, way too often. You can close the gaps as you work through the conversation, but most people don’t.

What’s up

The what’s up half doesn’t go much better. People jump straight into a watered-down version of what they want without ever really saying what’s going on, and the chances of doing high-level work with the model and actually solving your problem or taking advantage of your opportunity fall significantly.

The more the model understands about you and the problem you’re trying to solve, the better work it can do. The more someone treats it like a magic wand, assuming it will infer the situation or pull the details from the memory of some old conversation, the worse things go.

Underneath both

There’s something I run into all the time: you can’t solve a problem until you understand what the problem is. I talk with people about their problems regularly, and it’s very common that they have a hard time telling me what the problem actually is. That’s normal. Stating a problem well is genuinely hard.

The next two stops are the fix. A way of working that gets what’s up out of your head, sharpens what you want, and does both before you ask the AI for anything.

Everything here is how I actually work. Your reality may be different. Your company’s tools may not allow some of it, and you may work somewhere that talking out loud isn’t practical. I can’t know everybody’s circumstances, and this won’t fit every one of them. What I can do is teach the foundations underneath the way I work, so you can adapt them to your own reality with more confidence.

I use this method a lot. I used it to build this primer, as a matter of fact.

Three phases:

• Brain dump. I get everything out of my head and in front of the AI, and I ask for nothing.
• Q&A. The AI interviews me about what I left out.
• Refinement. I finally ask it for something, and we work it until it’s right.

Phase one: brain dump

When I start something real, I don’t ask the AI for anything. I dump. I get everything I can out of my head and in front of the AI, as messy as it comes. The problem I’m trying to solve, all the details I can think of, work I’ve already started, documents I have, a first attempt that didn’t go anywhere. Anything that describes the situation.

That rarely fits in one message. If I have a lot to get out of my head, I send it in pieces, so nothing sits in an input box long enough to get lost.

The only thing I allow during brain dump is an acknowledgment. One word, every time I send. My agents simply reply with “Word.” You can use whatever word you want. That way I can send as many times as I need to and it stays out of my way the whole time.

One thing that didn’t come naturally to me at first: working with AI at a higher level means challenging yourself to think about things in new ways. The biggest shift for me was realizing that sometimes the best way to move forward is to step back and resolve something at a higher level than the one where I’m struggling.

My favorite example is making a plan. If my sessions on a project keep going sideways, I can keep fighting at that level, or I can step back and spend a session producing a plan for the whole thing. That plan becomes the backbone context document for the entire project. Every session after it starts out already knowing the plan, so every session after it gets easier.

That’s the pattern: solve something earlier in the process than where the struggle is, and the work usually pays for itself twice, because you can keep using it. The purest version of this is stepping all the way back to the problem itself.

A session that only produces a clear problem is a session well spent

Sometimes the right use of the whole three-phase method is just to develop the problem thoroughly and “on paper.” Dump it, get questioned on it, revise it until it’s rock-solid, and end the session with a detailed problem statement.

What’s really the issue? What’s causing it? How do you wish it worked? No limitations, and no solving yet.

This process not only helps you truly understand something in your own head but also ensures the AI you are working with does too. Now it knows what it really needs to help you solve.

These tools lean toward agreeing with you. Tell one your plan and it will usually find something nice to say about it. So “the AI liked my idea” tells you almost nothing.

Ask for pushback on purpose. “What’s wrong with this plan?” “Argue against this.” You can even hand your idea to a second AI and ask it to attack what the first one produced.

They also guess with confidence

Sometimes an AI states something false as if it were plainly true. This is generally known as hallucination. Oftentimes models will answer incorrectly instead of admitting that they don’t know something. Other times they just make mistakes or get things wrong.

My take on handling it: expect these tools to strengthen your own understanding of what’s right, instead of expecting them to be right. That’s when hallucination becomes less of a problem.

And their training has a cutoff

A model’s training is limited. It usually has a cutoff point a number of months before today’s date, and the model doesn’t know what happened after it.

If your work needs current information, make sure web search is part of the picture. Most of the chat tools now search the web on their own unless that’s been turned off. It’s worth confirming yours actually did look something up instead of leaning on its training alone.

None of this makes them a fact machine

I don’t see these tools as fact machines. In time you learn what you can trust and what you would never trust. I see them as processors and amplifiers of my own ability more than answer machines.

They can produce answers and finished pieces of work, and there’s a time and a place for that. But to me, the biggest gains come from using them to leverage my own brain.

These tools cannot be responsible for anything. We can. We’re the ones who have to live with the results of our work, so we should be the ones making the decisions about what we’re going to do. And we already know from the last stop that they’re likely to make mistakes about things anyway. Why would we let them decide for us?

So don’t let what these things say decide what you’re going to do. Know your own ideals, know what you believe in, and let that be the driving force.

AI can help you organize your thinking and make a decision far better. You shouldn’t ever let it make the decision for you.

A question worth carrying

Here’s something I ask myself on a daily basis these days: does this support my values? I suggest you ask yourself the same question often, especially as you ramp up your exploration of these powerful tools.

When the answer is yes, keep going. When it’s no, stop and ask why you’re doing the thing at all. Don’t compromise those values.

I sometimes hear people say, “I don’t want to offload my thinking to AI.” Then don’t. Or, “I don’t use AI. I like writing.” Then write. If you love writing your emails, you should be the one writing your emails. That trap only catches people who aren’t in control and aren’t being deliberate. When your use of these tools is designed around your values, protecting the things you care about most and smoothing out the things you never wanted to be doing at all, you don’t end up saying, “I’ve offloaded my brain. I shouldn’t have done that.”

I think about artists here. I can’t imagine why an artist would want AI to make their art, unless they’re using it deliberately as part of the art itself. But I can absolutely imagine an artist using AI to smooth out everything that keeps them from putting a paintbrush against a canvas.

Knowing who you are, and making some commitments to yourself about how you’ll use these tools, is an important step. For getting good work done, and for keeping yourself out of trouble with them.

Almost everybody comes to AI thinking about efficiency. Doing what they already do more quickly and more easily. The first aha moments almost always arrive that way, and the gains are real. Massive, sometimes.

But the time you can save on what you already do is finite. Even if you could snap your fingers and have everything done, it’s still a limited amount of time and a limited amount of gain.

Capability is different. New capability has no upper limit. It opens new avenues, new opportunities, things you could never have done before.

About the job worry

Efficiency thinking is part of what fuels the worry about jobs. If everything gets faster, who’s still needed?

I’m not saying there’s nothing to worry about. I think some roles will go away, and many more will transform drastically. But a person who can wield AI at a high level becomes so much more capable that I don’t believe the natural response is “let’s not have this person on staff anymore.”

I think the question more companies will be asking is different: we can do new things now, so how do we take the people we have and do five or ten times what we were doing before? That’s why capability is the thing to build. It’s the strength.

“I don’t want to stop using my brain”

One more thing about the offloading-my-thinking worry from the last stop. My own experience has been the opposite of offloading. Since I learned to work this way, I’ve used my brain at a higher level than ever before, and it’s not close.

Working well with AI isn’t about tech-savviness. It’s about thinking things through, asking the right questions, and having the judgment and taste to know what you’re looking for, plus the confidence to say it.

Those are exactly the things a long career of solving problems builds: experience, judgment, instinct, and confidence, earned across a million different situations. If you’ve been around the block, you already have the most important skill in all of AI right now. It’s something I sometimes worry about for people just entering the workforce, who haven’t had a career’s worth of chances to develop it yet. You have it, and AI hands you a new place to spend it.

The energy principle

Something I discovered for myself along the way, and now deliberately chase: the work I hate most, or that’s hardest for me, costs me an inordinate amount of energy. The work I love, the work I’m best at, is almost the opposite. It gives energy back.

The better I’ve gotten at spending more time on the work that energizes me and less on the work that drains me, the happier, the more energized, and the more productive I’ve been. My ability to wield AI effectively is what makes that shift realistic to pursue.

Creative thinking matters more now

There used to be only so many things that were possible, and we all built our thinking inside that box. We’re expanding those limits now, to the point where I’m not sure there’s a box anymore. We don’t know where the edges of what we’re capable of are.

That makes creative thinking matter more than it ever has. It takes some bravery, because you have to be willing to try things that might not work. And it takes imagination, because you have to be able to uncover new things to try in the first place.

The gap is real

The distance between people who push into these tools and people who wait is widening. I’m not saying that to scare anybody. I’m saying it because it’s true.

The encouraging half is just as true: entering now, deliberately, still counts as early as far as I’m concerned. I talk to people about AI all the time, and trust me when I say the vast majority of people are not thinking about how to use this technology deliberately yet.

And the gains come fast. Think about joining a gym. At first, everything is new. You don’t know the form, you’re looking at a sheet of paper to see what you’re supposed to do, and you’re hunting around for the right machines. It takes time to get in the groove. It also hurts. You’re sore after the first day, and you might stay sore for weeks every time you go. Meanwhile, you don’t see any difference. It can take months to ever notice one.

Working with AI takes some getting used to as well. You have to commit to some new routines and do some things a little differently than you did them before. But it doesn’t hurt, and the gains can come so quickly that you don’t need months of commitment before you see a benefit from the work. A deliberate approach, understanding what you’re trying to do, and making some changes to the way you work can make an immediate impact and solve massive problems right away.

Leverage is anything you build into the way you work that keeps paying you back after you build it.

In the age of AI, I think it’s time to focus on leverage, because most of the things that create it are so much easier to do than they used to be. In the past, when you noticed something was off, you usually put it off, because fixing it wasn’t realistic. Now, especially when you’re working with agents, you can often implement the little change right when you think of it.

How easy depends on what you’re doing, but for most things, there’s now a much easier way to do them. And building things that pay you back on everything downstream is a compelling way to spend effort. The more leverage points you create, the more the benefits compound, the easier things get, and the more your focus goes where you want it instead of where you don’t.

The other half is paying attention to what slows you down. Blockers are negative leverage. They keep your gains from picking up steam, and they deserve more of your attention now than ever, because AI makes them easier to remove.

Writing things down

Documentation turns knowledge into leverage. A transcribed meeting turns a conversation into something you can hand to an AI later. A written-down process turns something that lived in one person’s head into something any of your tools can use. And this goes across way more things than capturing what the veterans know.

The reason comes straight from the context window: an AI only knows what got handed to it. Everything you get out of your head becomes something you can hand it.

A clean environment

Picture a computer environment that’s set up in a way you understand, designed for how you like to work. You know where everything is at all times, and you can flow through your machine doing the work you need to do.

Your time goes to thinking about the actual work. Not hunting for some file, or trying to remember where something is, or collecting pieces out of an inbox with fifty thousand emails in it.

We’ve all always known this matters. AI makes it easier to get there than it used to be, and the gain is bigger.

A home of your own for your context

There are ways to use these tools inside the products the companies make: projects, notebooks, settings. There are also ways to keep how you use AI more independent of any one company.

I think of it as a context library. A place of your own where your important context lives, in its own folder structure, outside any single system. When your context sits in its own library, any AI tool you have can use it. You get flexibility, and you’re not locked into one product. If you switch tools next year, your context comes with you.

Stacking gains

One of the great signs of fluency is using AI to make the other things you do in AI easier. Have the AI help you write the description of yourself that goes into your settings, and that description improves all the further work you do in that tool. Gains stack.

Making it easier on everybody else

High-level AI use also makes things easier for the people around you. If I need something from somebody, I try to give them something that makes it easy to give me what I need.

A dozen questions in an email means they have to open it, remember it, come back to it, and answer one question at a time, and there’s still a good chance something gets missed.

That really isn’t necessary anymore. If the situation allows it, I would usually just say, “Hey, can I have a couple of minutes of your time?” and record or transcribe that conversation, with their permission. I get the information I need, and I have the real-time opportunity to ask follow-up questions right then and there. A 15-minute conversation like this is almost definitely going to give me way more information than if they sat down and typed up all of their answers. Chances are it’ll be way easier on them too.

It’s pretty rare that AI does something 100% for you. Most of the time, AI gets you to 80% or so, real fast. And that 80% can look like 100% if you don’t look at it very closely.

If you turn in that 80% as finished work and somebody looks at it with scrutiny, it comes across as slop. Because it is. Shipped 80% is slop, no matter how polished it looked on the way out the door.

You can usually get closer to 100% by iterating. Doing the work with some consistency, smoothing out the wrinkles in your process, and making it a little better as you go. Sometimes that path takes you all the way. Either way, reviewing what comes out is your job, every time.

Stand by every letter

I think we, as responsible users of AI, need to dedicate ourselves to reviewing the work we do. To stand by every letter, every piece of code we ship, every decision we make. Stand by it as if AI was never involved, and be able to talk about it and understand what we’ve done.

That takes deliberation. The failure is saying “okay, that sounds good, let’s do it that way” without understanding what you just agreed to. If you don’t understand what you’re doing, spend a little time understanding it, so you have a reason for doing things a certain way instead of blindly trusting an AI.

In time, you learn what you can trust and what you wouldn’t ever trust.

When not to use AI

Fluency is partly knowing the difference between what AI can do and what it can’t. When to trust it, when to keep pushing, and when to reach for something simpler instead.

Sometimes AI is the wrong tool for the job, and a simpler answer wins. Even then, AI can still help. You can use it to think through, design, and build that simpler answer. The thing you end up running doesn’t have to have AI in it.

The better your fluency gets, the better you get at recognizing the right solution to a problem and judging whether AI is part of it. That judgment is what these foundations are for.

The big AI companies publish their own free courses, and they’re worth knowing about: Google AI Essentials, OpenAI Academy, and Anthropic’s AI Fluency course.

Ethan Mollick, a Wharton professor, wrote the book Co-Intelligence and publishes a newsletter called One Useful Thing. I read his book and quite like the way he thinks.

Those resources can show you the kinds of things the labs consider most important, and that can be a real part of how you understand this world. My experience is that the foundations of real practical, professional use, the things in this primer, aren’t touched much in those courses. That’s why this exists. They can still be useful. They just do a different job than this one does.

One podcast

I’m not selling anything here. I don’t know the guy, and I have no stake whatsoever in his show. If you want to keep your finger on the pulse of what’s happening in AI in a general, day-to-day way, subscribe to The AI Daily Brief with Nathaniel Whittemore. Nothing major will get by you. I find it an extremely helpful podcast for anybody trying to keep up with a world that feels impossible to keep up with.