Getting started with AI — Layer 2: what it already knows, and what you have to tell it
This is the second in the series. In Layer 1 we just had a conversation — got set up, tried a few real jobs, got a feel for it. This layer is about getting consistently good answers, and it all rests on one idea, so let's start there.
In Layer 1 I said it's a conversation, not a slot machine — if the first answer isn't right, just say so and it'll adjust. That's true. But there's one thing going on underneath it that, once you see it, explains most of what separates a good answer from a poor one.
Start with where its knowledge comes from. The AI was built by reading a huge amount of writing — books, articles, websites — up to a certain point in time, and all of that got baked in. That's why it can draft a decent email, explain a tax term, or suggest a recipe without you teaching it anything. It knows a great deal about the world.
But that baked-in knowledge has two hard edges.
First, it's general. It's the world's knowledge, not yours. Out of the box it knows nothing about your business, your Thursday meeting, your awkward client, or the letter on your desk — none of that was ever part of what it read.
Second, it's frozen. Its training stopped at a particular date — its "cut-off" — so it wasn't there for anything since, and it can't look things up on its own unless it's been connected to something that can. (More on that below.)
So here's the idea: every answer is a mix of two things — what it already knows in general, and what you put in front of it about your situation. It brings the general knowledge; you bring the specifics. It works from what it's got; it just doesn't have your half until you give it. Once you see that, a handful of things that used to feel random start to make sense.
1. Context — feed it your specifics
Most disappointing answers aren't the AI being dim. It has plenty of general knowledge to draw on — what it's missing is your half: the facts only you know. So it answers the question you did ask rather than the one you meant.
"Write me an email declining the meeting" gets you a generic brush-off. "Write a warm but firm email declining Thursday's meeting with my accountant — I want to move it to next week because I'm still waiting on figures from my bookkeeper" gets you something you can almost send as-is. Same tool, same skill with words. The difference is the specifics you put in front of it.
So the single most useful habit: give it the stuff. Don't describe the long email — paste it in. Don't summarise the report — hand it the report. Got a photo of a letter, a spreadsheet, a menu, a form? Drop it in and ask your question about that. (Layer 1's "here's a long document, give me the three key points" was the first taste of this — the same move, all the time.)
A rule of thumb: before you send, ask "have I given it the specifics a knowledgeable stranger would need to do this well?" The general knowledge it already has. The specifics are your job.
2. Context-rot — why a long chat goes downhill
Here's one almost nobody warns you about. You start a chat, it's sharp, you keep going… and forty messages later it's gone vague, forgotten a detail you gave it early on, or started going in circles. You didn't do anything wrong. The conversation just got too full.
The specifics you hand it live in the conversation itself — and in a long chat that becomes a sprawling transcript. The important thing you said at the top is now buried under a hundred later messages, and it gets crowded out. People call this context-rot: the longer and more meandering the chat, the more the signal gets lost in the noise.
The fix is low-tech: start a fresh chat. When a conversation has wandered, or you're moving to a new task, open a new one and re-state the handful of facts that matter. It feels like starting over, but you're handing it a clean desk instead of a cluttered one, and the answers improve. One good conversation per job beats one endless conversation for everything.
3. Memory — when it carries your specifics over for you
Now a feature that does that second half for you. You may have noticed the AI occasionally remembering something about you between chats — your name, that you run a small business, that you prefer short answers. The newer apps have a memory feature: a small, separate note the tool keeps about you and puts in front of it at the start of each new conversation.
This isn't a contradiction of anything above — it's the same principle, automated. Memory is the tool doing the "bring the specifics" job on your behalf, so you don't re-explain who you are every time. It's also the thin end of a bigger wedge: these tools can increasingly be connected to your own sources — your files, your inbox — so they pull in the specifics themselves rather than wait for you to paste them. You don't need to set any of that up today; just know that "getting the right stuff in front of it" is a job that can be handed off. (That's much of what a later layer is about.)
Memory is useful, but worth keeping half an eye on: if it once noted something that's no longer true, that stale fact rides along into every chat. If answers start feeling off-target, it's worth a look. In Settings → Personalisation (or "Memory") you can see what it's remembered and clear anything out of date.
4. Up-to-date facts — it has to look them up
Remember the second hard edge: its knowledge is frozen at its training cut-off. Ask it about last week's news, today's weather, a live price, or this season's rules, and it's working from a world that stopped some months ago. Left alone it'll either say it can't be sure, or answer confidently from the stale picture it does have.
The fix is web search — letting it read the live internet before it answers, so it's grounded in what's true now rather than what was true at its cut-off. When it searches, it'll usually say so and cite the pages it used, so you can click through and check.
Both of the big tools can do this, but they behave a little differently, and it's worth knowing which you're using:
- ChatGPT — search is on for everyone by default. It looks things up on its own whenever a question would benefit from current information, so most of the time you don't have to do anything. If you want to force it, open the tools menu (the "+" beside the message box), choose Search, and ask your question.
- Claude — search is off until you switch it on, and it's a one-time job: click the tools icon in the message box, find Web search in the list, and turn the toggle on. After that Claude searches by itself when a question needs it, the same as ChatGPT.
A simple tell either way: if an answer about anything current comes back with no mention of searching and no links, treat it as the frozen version — just say "search the web and check."
5. …and this is also why it sometimes makes things up
I promised at the end of Layer 1 we'd cover knowing when to trust it. This is that — and there's an apparent puzzle to clear up first.
If it was trained on real knowledge, why would it ever make things up? The answer is in how that knowledge is held. It didn't file away a set of facts to look up later, like an encyclopedia. As Layer 1 put it, at heart it's a prediction machine: it learned the patterns in everything it read, and it answers by predicting, a piece at a time, what a good response would look like.
That prediction is the point — it's what makes the tool useful rather than a search box. Because it works from patterns rather than stored records, it can go past what it literally read: apply a general principle to your specific case, combine two ideas, phrase something that appears nowhere in its training. It isn't only returning knowledge; it's generating from it.
The catch is that the same machinery runs whether or not it has anything solid to go on. When you ask for something in neither half — not well covered by what it absorbed in training, and not in what you gave it — it doesn't stop and say "I don't know". It does what it always does: predicts the most plausible-looking continuation. And a plausible-looking answer isn't always a true one. So you get a made-up statistic, a book that doesn't exist, a figure stated with the same confidence as a real one. There's no internal gauge that swings to "unsure" — fluent and correct look the same on the way out. People call this a hallucination, and it's the one thing to keep a hand on the wheel for.
So it follows from the same two halves when to be on guard:
- When it's working from what you gave it — summarise this, reword that, pull the numbers out of this document — it's on solid ground. Trust it.
- When it's drawing on well-trodden general knowledge — explaining a common idea, drafting a standard letter — it's usually reliable; that's the pattern-matching at its best.
- When it's reaching for a specific it wasn't given and didn't have well-covered in training — precise numbers, names, dates, quotes, anything recent (past its cut-off), legal or medical specifics — that's where it's predicting into a gap, and where it can bluff. Check it.
The habit: for anything that matters, ask it "how do you know that?" or check the key fact yourself. Often, giving it the missing specifics up front is the fix — with no grounding it predicted; hand it the source and it has something real to work from.
That's Layer 2
One idea, several payoffs. It comes knowing a great deal in general, but nothing about your world until you tell it. Feed it your specifics, keep your chats focused, know what its memory is doing, switch on search when you need today's facts, and stay alert for the moments it's predicting into a gap. None of it is technical — it's just knowing what the thing in front of you is actually working with.
Do that and you'll get better answers than most people, from the same free tools. Next layer, we go further — connecting AI to your own documents and tools properly, so it's not just working from what you paste into a chat, but from your stuff, on tap.
I'm Simon — I run Simon Studios, an AI-first development studio in Teddington. I help small businesses and individuals work out where AI genuinely helps (and where it doesn't), then build it — no jargon, no hype. A lot of what I build is exactly this idea taken further: AI wired into a business's own information, so it always has the right context in front of it and stops guessing. If that sounds useful, that's the conversation I love to have. Book a free intro call at cal.com/simonstudios/free-intro-call and we'll work out whether there's something real here for you.