Understanding your AI assistant

You’ve been working with Claude for a few modules now without needing to know much about it — which was the point. But a little understanding of what’s actually happening pays off fast: it helps you work smarter, get better results, and avoid quietly burning through what you’re paying for. Nothing here is technical. Think of it as learning what the dials in your car do, after you’ve already learned to drive.

What the assistant actually is

Claude is an AI model A computer program trained on enormous amounts of text so it can predict useful, sensible responses. Not a person, and not a search engine. — a program trained on a huge amount of writing and code until it got very good at producing sensible, helpful responses. A few things that are genuinely useful to keep in mind:

• It isn’t a person and it doesn’t “know” you. It works entirely from the conversation in front of it. That’s why being clear and specific helps so much.
• It can be confidently wrong. Most of the time it’s remarkably good, but it can occasionally state something untrue as if it’s certain. For the building you’ll do, you catch this naturally by looking at the result — if the page is wrong, you just say so.
• It works best in a back-and-forth. You’re not writing one perfect instruction; you’re having a conversation, nudging toward what you want.

Tokens: the thing you’re actually spending

Here’s the one bit of plumbing worth understanding, because everything else hangs off it. The assistant doesn’t read whole words — it chops all text into little pieces called tokens The small chunks of text an AI reads and writes. Roughly speaking, a token is about three-quarters of a word. Everything you send and everything Claude replies counts as tokens. . A token is roughly three-quarters of a word, so a short sentence might be a dozen tokens.

Why you should care: tokens are the unit your plan is measured in. Every message you send, every file Claude reads, and every word it writes back is counted in tokens — and your allowance is really just a big bucket of tokens that refills over time.

That single fact explains almost every habit in this module:

• A longer conversation has more tokens on the table, so each new reply costs more to produce.
• A bigger, more powerful model charges more tokens for the same work.
• Asking it to read your whole project when it only needed one file spends tokens you didn’t have to.

None of this means counting tokens yourself — you never will. It just means a few light habits (which model, how long a conversation, how big a request) quietly decide whether your plan feels generous or cramped. The rest of this module is those habits.

There’s more than one model

“Claude” is actually a small family of models, and they trade off power against speed and cost. The names change over time, but they run from small and cheap up to a powerful, pricey frontier model:

• Sonnet — the all-rounder, and the one you should live in. It handles the large majority of everyday building beautifully while sipping your allowance. Think of it as your default — the setting you build on day to day.
• Opus — a big step up in power and thoroughness, and the thirstiest of the everyday models by far. It can produce noticeably better work on genuinely hard problems, but it burns through your allowance shockingly fast — many times quicker than Sonnet for the same conversation. Reach for it only when Sonnet is truly stuck, then switch back.
• Fable — the newest and most capable of all, Anthropic’s frontier model, sitting a tier above Opus. It’s built for big, professional jobs — think an expert agent running for hours across an entire codebase — and it’s priced to match: far and away the most expensive, several times thirstier again than Opus. For the building you’ll do here it’s overkill, and it would drain a beginner’s allowance alarmingly fast. Good to know it exists; leave it be for now.
• Haiku — the small, fast, cheap one. Fine to dip a toe into for trivial jobs (a quick rename, a one-line question) where you want to spend almost nothing. Be warned, though: it’s noticeably less capable, so it’ll struggle with the real building work. Handy to know exists; not where you’ll spend your time.

So even though there are four, there are really just two you’ll think about: Sonnet for almost everything, and Opus for the rare hard problem. The fancy frontier model (Fable) and the tiny one (Haiku) sit at the edges and aren’t part of your day-to-day. That’s the whole decision.

Check which model you’re on, and switch it

Inside Claude Code (after you’ve run claude), you can ask which model you’re using:

/status

To switch to the efficient everyday model, type:

/model sonnet

Do this at the start of every session, before you ask for anything else. Claude Code remembers your choice for next time, so once you’ve set Sonnet it should stick between sessions — but it costs nothing to check with /status and switch again if it’s drifted back. The rule of thumb is simple: stay on Sonnet by default, and only switch up to Opus (/model opus) when you hit a genuinely hard problem Sonnet keeps failing — then switch straight back to Sonnet (/model sonnet) afterward. Most people build entire projects without ever leaving Sonnet.

Effort: how hard it thinks

Newer versions of Claude Code can also vary how much thinking they do before answering — sometimes called effort. You can nudge this just by how you ask: words like “think carefully about this” or “take your time” tell it to reason more deeply, while a plain request keeps things quick.

The trade-off is the same one as everything else here — it’s measured in tokens:

• More thinking can produce better answers on tricky, fiddly problems… but it uses more tokens, so it spends your allowance faster.
• Less thinking is quicker and cheaper, and it’s perfectly good for the everyday building you’ll mostly be doing.

You don’t need to fiddle with any setting. Just know the lever exists: for routine work, let it answer normally; when you’re truly stuck on something gnarly, then say “think carefully about this one.” Using deep thinking on every little request is a quiet way to drain your allowance for no real benefit.

The plans, and whether you’ll ever need more

You’re already on Pro — you set it up back in Module 2. This section is just so you understand what that gives you, and how to tell if you’d ever want more. (The free tier, for contrast, is fine for casual chatting on the website but won’t run Claude Code at all — which is why Pro was step one.) The key thing to know: a higher plan doesn’t make the assistant smarter Anthropic's paid plans for Claude. A higher tier gives a bigger usage allowance — more room to work — not a smarter assistant. ; it just gives you a bigger allowance — more room to work before you hit a limit.

• Pro — about $24 CAD/month. This is where you start. It unlocks Claude Code, Sonnet, and the website. For learning and small projects it’s the right first step, and it’s the cheapest way in.
• Max — from about $140 CAD/month. Several times Pro’s allowance (it comes in a 5× tier and a larger 20× tier). The upgrade to consider only if you find you genuinely love this and keep bumping into Pro’s limits. There’s no pretending here — this is a real expense, the cost of a serious hobby. Don’t jump to it on day one.

If you ever want to change plans

It’s the same screen where you started: go to claude.ai↗ (opens in a new tab) , click your name or initials in the bottom-left corner, open Settings → Billing, and you can move up to Max (or back down) anytime. But honestly — give Pro and the good habits in this module a real try first. Most people learning never need more.

Your usage allowance (and how not to blow through it)

Your paid plan doesn’t give you unlimited AI — it gives you a generous allowance that refills on a schedule. There are two timers running at once:

• A 5-hour window A usage limit that starts the moment you send your first message and resets about five hours later. : it begins with your first message and resets roughly five hours after. Heavy use in a short burst is what fills this one.
• A weekly cap: a larger budget that refills every seven days. Even if your 5-hour windows keep resetting, very heavy use can fill the weekly bucket and ask you to wait until it resets.

See exactly where you stand

When you want the real numbers — how much you’ve used and when it resets — there’s one authoritative place:

• Go to claude.ai↗ (opens in a new tab) , open Settings, then Usage. This counts everything, across the website and every computer you run Claude Code on.

And for a quick check of what’s filling up your current conversation, type this to Claude:

/context

Context: the assistant’s short-term memory

Every conversation has a context window The amount of conversation an AI can actively hold at once — its working memory for the current session. — think of it as the assistant’s desk. It can only keep so much on the desk at once: everything you’ve said, every file it’s read, every result it’s seen this session.

Here’s why this matters for you, not just the computer: a fuller desk doesn’t only cost more tokens (remember — everything on the desk gets re-read on every reply), it also tends to give you worse answers. When the desk is cluttered with three unrelated tasks and a dozen files from an hour ago, the assistant is more easily distracted and more likely to lose the thread. A light, focused desk is both cheaper and smarter. Keeping your context light is one of the highest-value habits you can build.

When the desk gets full, Claude Code automatically tidies it — it compacts When the assistant summarizes the earlier parts of a long conversation to make room to keep going. the conversation, summarizing the older parts down to a short recap so it can keep going. You might see a brief “compacting…” message. That’s completely normal — it’s housekeeping, not a problem, and your files are never touched. Just know that a summary is necessarily lossy: little details from early on can get smoothed over, so for a genuinely new task you’re better off starting clean than relying on a compact.

Two commands give you control over the desk:

/clear

/compact

The single best habit here: start a fresh conversation for each new task. When you finish one thing and move to something unrelated, type /clear. A focused, uncluttered conversation gives better results and uses less of your allowance. Don’t be precious about it — clearing often is a sign you’re working well, not a waste.

Give your project a memory

There’s an obvious catch in everything above: if a light, focused conversation is best, and /clear wipes the slate between tasks — then how does Claude remember what your project is and where you left off? The answer is one of the most useful habits in this whole guide: don’t keep the project’s memory in the conversation, keep it in a file inside the project folder.

The trick is to have Claude write and maintain a couple of plain notes files that live with your project (and get backed up to GitHub along with everything else):

• A PLAN.md — what you’re building, the steps still to do, and what’s done so far. A living to-do list for the project.
• A NOTES.md (or README.md, which you’ll meet in Module 5) — decisions you’ve made, how things are set up, anything you’d otherwise have to re-explain.

Because these are real files, they survive /clear, they survive closing your laptop, and they survive next week. When you sit back down, you start a fresh conversation (cheap, focused) and simply point Claude at the plan:

Read PLAN.md to see where we left off, then let's continue with
the next unchecked step.

To set this up the first time, just ask:

Create a PLAN.md in this project. Write down what we're building,
break it into small steps as a checklist, and mark what's already
done. Update it as we go so I can always pick up where I left off.

There’s a second, related file worth knowing about — and it does a different job, so don’t mix them up. It’s called CLAUDE.md A small notes file Claude reads automatically at the very start of every conversation. The place for standing rules and facts about your project — kept short on purpose. , and the special thing about it is that Claude reads it automatically at the start of every conversation — you never have to say “read this.” That makes it the perfect home for your project’s standing rules: what the project is, how to run it, and any “always do this, never do that” preferences you want respected every time.

Keep the two straight, because they pull in opposite directions:

• CLAUDE.md = the rules of the workshop. Small and stable. Because it’s loaded into every conversation, every line costs a little allowance every time — so it should stay short, and it is not the place to keep your plans or to-do lists.
• PLAN.md = today’s job sheet. Allowed to grow and change; read only when you point Claude at it. This is where the actual planning lives.

To set up the rules file, just ask (or run the built-in /init command, which does the same thing):

Set up a CLAUDE.md for this project with a short summary of what it
is and how to run it. Keep it brief — just the standing facts and
rules you should know every session. Don't put our task plan in it.

The habits that keep it cheap and smooth