The words nobody explained to you

Nine issues of vocabulary I never explained

I've used a lot of terms in this newsletter without ever stopping to define them. Projects, context, custom instructions, cutoff dates. If you joined somewhere in the middle, you've probably been reading around words that nobody ever taught you, and that is on me.

The bigger problem is that some of the most common terms in this space don't mean what you'd assume they mean, and that ends up costing you more than the unfamiliar ones do. Words like memory and learns sound like they mean one thing and actually describe something else, which leads coaches to either trust the tool more than they should or give up on something that would have worked fine if they had set it up differently.

So here are the terms worth knowing, split into two groups. The ones that change what you get back, and the ones that quietly mislead you.

The terms that change what you get

• Context is everything the tool can see at the moment you ask your question. Your message, the conversation up to that point, any files you attached, anything sitting in the project, and anything memory pulled in on its own. If it is not in there, the tool has no way to know it, no matter how obvious it seems to you, which is why almost every bad output traces back to context you assumed it had.
• Context window is how much it can hold at once. Every conversation has a ceiling, and once you go past it, the earliest parts of that conversation stop being visible to the tool. This is why a chat you have been running for months starts feeling like it lost track of things. It did.
• Project is a workspace that holds instructions and files that apply to every conversation inside it, which is the fix for retyping the same background information every time you open a new chat.
• Custom instructions are standing orders that apply to every chat in a project, or across your whole account. This is where you tell it how you want things done once instead of asking for the same thing over and over.
• Cutoff date is the point where its knowledge of the world stops. Anything after that, it either does not know or has to go look up, so it is worth knowing before you ask about anything recent.

The terms that mislead you

• Memory is the one people get wrong in both directions. Claude and ChatGPT both save facts about you as you work, and on Claude that turned on for every account back in March. It is not saving your conversations word for word. It pulls out things that seem worth keeping, your role, how you like information formatted, what you are working on, and puts those back in front of the model in later chats. So it can absolutely know something in a new conversation that you told it last week. What it does not have is the full detail of what you actually said, which is why it will sometimes surprise you by remembering something and other times seem to have missed the point entirely.
• Learns is where people go wrong. The model itself does not learn from you. When you correct something and it adjusts, the model has not changed and it will not be any better at that thing for the next coach who asks. What can happen is that the system saves your correction as a fact and hands it back to the model later, which looks like learning from the outside but works completely differently underneath. The practical version is this. If it matters, do not count on it being picked up. Put it in a project or your custom instructions where you know it will be there.
• Training gets used two different ways. Most of the time it refers to the process that built the model, which happened long before you opened it. It also comes up when people talk about what these companies do with the things you type into them, and that version has real consequences for anyone handling athlete information. That is next week's issue, because it deserves more attention.
• Thinking is what some tools call it when they work through a problem in steps before answering, and it genuinely does produce better results on complicated questions. It is not deliberation the way you deliberate on the drive home. It is the model reasoning in text before it commits to an answer, which catches errors a fast response would miss. Worth turning on when the question is hard.
• Hallucination is the industry's word for confidently making something up. It is a soft word for a real problem, and it is the reason Issue #8 spent the whole issue on where to check.

Where I think this actually matters

You don't need to memorize any of this, because there is really one idea underneath all of it.

The tool only ever knows what is in front of it when you hit send. Inside a chat that is the whole conversation, which is why you can come back days later and pick up where you left off. Outside of it, you are relying on whatever the system happened to save, or on what you deliberately put in a project. Memory helps, but it is selective and you do not control what it keeps. Everything I've written about in this newsletter, the Program Card, projects, custom instructions, saving your phase briefs as documents, exists so that the important stuff is there on purpose instead of by luck.

So when an answer comes back off, the way I think about it is to run three questions. Did it actually have what it needed, or did I assume it knew something I never told it? Has this conversation been going long enough that the early part dropped out? Is it working from something it saved about me months ago that is not true anymore? That last one catches more than you'd expect, and it is the same problem we talked about with keeping your Program Card current.

Where the coach stays in charge

The vocabulary in this space changes fast and half these terms will get renamed or replaced in a couple of years. The mechanics underneath them move a lot slower, so what actually matters is that you understand what the tool can see and what it cannot. Everything else is labeling.

The Coach's Scouting Report

• Signal. The tool answers based on what is in front of it when you hit send. Memory fills some of that in on its own, but what you put there on purpose is the part you can count on.
• Noise. Assuming the model is learning from your corrections. Trusting a saved fact about your program that stopped being true two seasons ago.
• Next Rep. Go into your settings and look at what it has saved about you. Most coaches have never opened that page. Delete anything that is out of date, and take the things that actually matter and put them in a project where you know they will be used.

See you next Sunday.