The Weight of Words
What AI taught me about the cost of saying too much.
7 minute read
I used to think words were free.
In conversation, they are. You can ramble. You can repeat yourself. You can start a sentence, abandon it, start over, circle back, say "you know what I mean" four times, and the person sitting across from you will nod and follow along because human brains are extraordinary at filtering noise and finding signal. We've been doing it for thousands of years. We're built for it.
I spent the first forty-something years of my life talking the way most people talk. I said what I needed to say, plus a little extra, plus some filler, plus a few tangents, plus whatever else came to mind. And it worked. It worked because humans are generous listeners. They fill in gaps. They interpret. They hear what you meant even when it's buried under what you actually said.
And then I started working with AI every day. And I learned something I can't unlearn.
Words have weight. Real, measurable, consequential weight. And every word I spend on noise is a word I can't spend on signal.
What a Token Actually Is
If you've used AI for any amount of time, you've probably heard the word "token" thrown around. It sounds technical. It sounds like something you don't need to worry about. It sounds like a backend detail meant for engineers.
It isn't. Tokens are the single most important concept in understanding how AI works, how it thinks, and why it sometimes stops thinking well. And you don't need a computer science degree to understand them. You just need a good metaphor.
A token is a piece of a word. Not a whole word. Not a letter. Something in between. The word "understand" is two tokens. The word "the" is one token. A long, complicated word might be three or four. Punctuation counts. Spaces count. Everything you type gets broken into these small pieces, and every piece takes up space.
Think of it like packing a suitcase.
You're going on a trip. You have one suitcase. It holds a fixed amount. Every shirt, every pair of shoes, every charger and toiletry bag takes up space. If you pack smart, you fit everything you need and you're prepared for whatever the trip throws at you. If you pack carelessly, throwing in things you don't need, duplicates, stuff "just in case," you run out of room before you've packed what actually matters.
Tokens are the contents of your suitcase. The suitcase itself is called the context window. And the context window is the total amount of information your AI can hold in its head at one time.
The Suitcase Isn't Empty When You Open It
Here's the part that surprises people.
You might assume that when you start a fresh conversation, you're working with a completely empty suitcase. You're not. Before you type a single word, your AI has already loaded its essentials. Its system instructions. Its skills. Its tool definitions. If you're working in something like Claude Code, your CLAUDE.md file, your project configuration, your custom skills, your MCP server connections, all of that gets packed into the suitcase before you even say hello.
Think of it like the stuff that's already in your luggage before you start packing for a trip. The toiletry bag that lives in the suitcase. The charger you never unpack. The travel pillow tucked in the corner. Useful stuff. Essential stuff. But it takes up space. And if you've never looked at how much space those essentials are consuming, you might not realize how much room you actually have left for the trip itself.
This is why reviewing your foundational files matters. Your CLAUDE.md, your skills, your global instructions, these are the permanent residents of your context window. If they're bloated, if they contain instructions you don't need anymore or context that's outdated, they're eating into your working space every single session. I go back and review mine regularly. I trim what's stale. I tighten what's verbose. I make sure the things that live in my suitcase permanently are earning their space.
Because the room you have to work in isn't the full context window. It's the context window minus everything that was already there when you arrived.
The Room Gets Smaller
Every conversation starts with whatever space is left after the essentials are loaded. Your first message is clean. The AI reads it, processes it, and responds with full access to everything you've said. It's sharp. It's focused. It remembers your instructions. It builds on your context. The room feels big and the AI has space to think.
But the room doesn't stay big.
Every message you send adds tokens. Every response the AI gives back adds tokens. Your instructions, your context, your follow-up questions, the AI's answers, all of it accumulates. The suitcase fills. And unlike a real suitcase, you can't see it filling. There's no progress bar. There's no warning light. You just keep talking, and the AI keeps responding, and somewhere around mid-afternoon the responses start getting a little less sharp, a little less specific, a little less connected to what you said that morning.
That's not the AI getting tired. That's the room getting smaller.
When the context window fills up, the AI has to make choices. It starts letting go of the earliest parts of your conversation to make room for the newest parts. Your carefully crafted initial instructions, the context you set up at the beginning, the preferences you stated, the scope you defined, all of it starts sliding off the edge. And the AI doesn't tell you it's happening. It just quietly loses pieces of your conversation while continuing to respond as if everything is fine.
This is why your brilliant morning session feels dumb by afternoon. This is why the AI starts repeating itself, or forgets something you told it three hours ago, or suddenly ignores a constraint you set at the beginning. You didn't break it. You filled it.
I Learned This the Hard Way
I run AI sessions all day. I build systems, write code, solve problems, create documents, all of it inside Claude Code. And there was a period where I couldn't figure out why my sessions kept degrading. I'd start strong, the AI would be sharp and responsive, and then hours later I'd be fighting it. Repeating instructions. Getting responses that missed the point. Wondering why a tool that was brilliant at 9 AM was struggling at 2 PM.
I was blaming the tool. The tool was fine. I was filling the suitcase.
I was pasting in long documents without trimming them. I was writing paragraphs where sentences would have done. I was repeating context the AI already had because I wasn't sure it remembered, which ironically pushed out the things it actually needed to remember. I was treating the context window like it was infinite because it felt infinite. It isn't.
The day I understood tokens was the day my entire workflow changed. I started being deliberate about what I put in. I started trimming. I started asking myself, before every message: does the AI need this? Does this add signal, or is this just noise I'm comfortable with? I started fresh sessions instead of pushing exhausted ones. I started treating every word like it cost something.
Because it does.
Words Cost More Than You Think
If you're using a subscription, you might not think about cost at all. You pay your $20 a month and you type whatever you want and the responses come back and life goes on.
But even on a subscription, tokens matter. Your plan has limits. Those limits are measured in tokens. Every bloated prompt, every unnecessarily long paste, every conversation you push past its useful life is spending capacity you could be using on something else. And if you're on the API, building tools, running agents, the cost is literal. Tokens have a price per thousand. The more you use, the more you pay.
I've watched token costs balloon on projects where the prompts were sloppy. I've seen the difference between a well-structured 200-token prompt that gets a perfect result and a rambling 2,000-token prompt that gets something mediocre. Ten times the words. Worse outcome. Because the AI had to wade through nine hundred tokens of noise to find the hundred tokens of signal.
More words is not more helpful. More words is more weight.
Signal and Noise
Here's the part of this that changed me, and not just as someone who works with AI.
Working with AI has made me a fundamentally better communicator. In every context. Not just in prompts. In emails. In meetings. In conversations with my kids. In the way I explain things to friends who ask me what I do all day.
Because AI taught me something that no class, no certification, no twenty years of professional experience ever taught me with this kind of clarity: every word you say either adds signal or adds noise. There is no neutral. Every sentence you speak is either moving the conversation forward or filling the room with furniture that makes it harder to move.
When I write a prompt now, I weigh every word. Not because I'm trying to be clever. Because I've seen what happens when I don't. I've seen the Roomba power off at my door because I said "never come in my office" when I meant "skip my office on your route." I've seen a card image take over my entire screen because I said "large and visible" without defining what that meant. I've spent three hours debugging a business rule because I pasted the same instruction over and over, filling the context window with repetition instead of stepping back and asking a better question.
Every one of those failures was a weight problem. Too many words. Wrong words. Words that said something I didn't mean, or said nothing at all but took up space anyway.
The Discipline of Less
I'm not saying be terse. I'm not saying strip every conversation to its bones and talk to your AI like you're sending a telegram. Context matters. Detail matters. The AI needs enough information to understand what you're asking and why.
But there is a discipline to this that most people haven't learned yet because most people don't know the suitcase exists.
When I sit down to write a prompt now, I think about three things. What does the AI need to know? What does the AI need to do? And what can I leave out? That third question is the one that changed everything. Because the answer is almost always: more than I think.
You don't need to explain why you need something unless the why changes the output. You don't need to provide your entire document when the AI only needs two paragraphs. You don't need to restate your instructions every fourth message because you're worried the AI forgot. If the AI forgot, it's because the room is full. Adding more words to a full room doesn't help. Starting a new room does.
This is the unglamorous part of AI literacy. It's not about knowing which model to use or how to write a clever prompt. It's about understanding that the space your AI has to think is finite, and the way you use that space determines everything about the quality of what you get back.
How I Pack the Suitcase
I didn't figure this out overnight. It took months of hitting walls, watching sessions degrade, and slowly building a system that respects the reality of how these tools actually work. But I want to show you what I landed on, because it changed everything.
Every day starts clean. Fresh sessions. No carryover from yesterday's conversations. I learned the hard way that picking up where you left off sounds efficient but it's the fastest way to fill your suitcase with yesterday's luggage.
Before I touch anything, I run what I call a daily start. I set my goals for the day. I try to keep it under ten. Each of my sessions claims a goal, and it knows that goal is its focus. Not everything. Not whatever comes up. One goal. That's its job today.
Memories from yesterday get read in, but not the whole conversation. I let Claude manage those memories, and I've asked it to keep them concise. Take up as few tokens as possible. Carry the signal. Leave the noise behind. This is the part that took me the longest to trust, letting the AI decide what's worth remembering. But it turns out that when you ask it to be efficient with your context, it's remarkably good at it.
During the day, I pay attention. I've asked Claude to watch its own context window and warn me when a session is getting long. Whenever an AI tells me we've been going for a while, I take the time to stop and pre-compact. I built custom skills for this. The AI saves the important parts of the conversation to memory, the decisions, the progress, the things the next session would need to know, and we let go of everything else. It's like repacking your suitcase in the middle of the trip. You keep what matters. You leave the rest at the hotel.
Every session ends with a wrap. Each session reports to a markdown document: what it did that day, what progress it made on its goal, anything it learned that the next session would benefit from. When tomorrow's session picks up that goal, it doesn't need the full conversation history. It needs a day bag. Just the essentials. Just what's relevant to today.
I used to let my sessions run and run because I thought history mattered. I thought more context meant better results. I thought the AI needed to remember everything to understand anything. I was wrong. What I learned is that carrying less into each session actually makes Claude work better. A day bag, not a suitcase full of last week's clothes. The AI thinks more clearly when it has room to think.
The Lever Gets Longer
I started writing this post before Anthropic released Fable 5. And I almost set it aside, because I wondered if a model this powerful made the whole argument obsolete. These new models are remarkable at understanding intent. They catch nuance that older models missed. They fill gaps that used to require explicit instruction. Maybe, I thought, the weight of words matters less now. Maybe the model is smart enough to carry the load for both of us.
I had it backwards.
The more powerful the model, the more weight every word carries. Not less. Think about what a lever does. A longer lever means a small movement on your end produces a much bigger movement on the other end. That's what these models are. Levers. And they just got dramatically longer.
When I give Fable 5 a well-aimed sentence, it launches an amount of correct work that would have been unthinkable a year ago. Entire systems. Complete documents. Multi-step solutions that used to take a full day of back and forth. One sentence, enormous output.
And when I give it a sloppy sentence, it launches an enormous amount of wrong work. Faster, more thoroughly, and more confidently than any model before it. The Roomba problem at scale. My vacuum powered off at a door. A careless instruction to a model this capable doesn't power off a vacuum. It builds the wrong system. Beautifully. Completely. Exactly to the specification I didn't mean to give.
Smarter models don't make your words matter less. They amplify whatever you hand them. The lever got longer. That's not a reason to relax your grip.
It's a reason to aim.
The Scale Changed Me
I weigh my words now. All of them. Not just the ones I type into Claude.
My boss said something to me recently that stopped me. He mentioned, gently, that I'm quieter than he remembers me being. He's brought it up a few times now, in that subtle way people do when they've noticed something but aren't sure what it means. I don't talk much in meetings anymore. He feels like he does all the talking.
He's right. And it took me a while to understand why.
It ties back to this. Words have weight. They have meaning. And I am less likely to waste them than I used to be. I spend more time thinking now, processing the world around me, than I do verbalizing it. When I do speak, I've usually already done the weighing. I know what I want to say and why it's worth the space it takes up. That's not me disengaging. That's me packing the suitcase before I open my mouth.
It's a strange thing to realize that a technology changed your personality. Not your opinions. Not your habits. Your actual behavior in a room full of people. But it did. I am quieter because I've learned what words cost. And once you know what they cost, you stop spending them carelessly.
I catch myself rambling in an email and I stop. I hear myself repeating a point in a meeting and I pull back. I notice when I'm explaining something with ten sentences that could be said in three, and I hear this voice in the back of my head that says: you're filling the suitcase.
AI didn't teach me to communicate better because it critiqued my style. It taught me because it put a scale under every word I said and showed me the reading in real time. When every sentence either helps or hurts. When the space is finite and the only person responsible for what's in it is you.
Most technology makes us lazier with language. Texting made us shorter. Email made us sloppier. Social media made us louder. AI is the first technology I've ever used that made me more precise. More intentional. More aware of the distance between what I said and what I meant.
And that, more than any tool or technique or prompt template, is the thing that made me better at this. Not better at AI. Better at communicating. Better at thinking. Better at knowing what I actually want to say before I open my mouth.
Words are not free. They never were. We just didn't have a system that counted them before.
Now we do.
Dacia writes about AI for real people at Speak Human. If you're trying to figure out how to actually use these tools in your everyday life, you're in the right place.