Notes: Know What It Costs and Keep Long Sessions on Track

In This Section, You Will Master:

• Reading what a session costs, and what that number does and does not mean
• The two settings that quietly decide how much of your day a piece of work takes
• Matching effort and model to the stakes, rather than reaching for the strongest by default
• Catching session drift before it costs you, and handing off cleanly
• A setup you chose on purpose, with a reason written next to every choice

How to Use These Notes:

These notes are designed as quick-reference companions to your video learning.

• Use them to review key concepts after watching lectures
• Reference them when you need a quick reminder of frameworks and tools
• 🎯 Core Concepts = Essential terms and frameworks
• 🚀 Actionable Items = Practical implementation steps

The Growth Payoff:

You stop hoping the defaults were fine. You can say what your work costs, why the model and the mode sit where you put them, and how to keep a long session honest. Knowing what your work costs, and being able to say what it is worth, is half of what makes someone trusted with real budgets.

---

Lecture: What the meter is for (Tokens, Cost and Context)

Something decided how much of your day that piece of work was going to take, and it was not you.

🎯 Core Concepts:

• Two settings did most of the deciding: how hard Claude thinks before it answers, and which of Claude's brains does the answering
• Both have been sitting wherever they came set since the day you installed this, and neither has ever come up
• The quieter half of this section is drift, where a session starts arguing for something you had both settled an hour earlier
• What you walk out holding is small and all of it is yours: a model, a resting mode, one more line in the file guarding your key, and a short page saying why

---

Lecture: Will it cost me $5 or $500? (Understanding Tokens)

The question that is hard to type out loud, taken apart starting from your own number rather than from a definition.

🎯 Core Concepts:

• On a subscription, that figure is not an extra bill on top of what you already pay. It is what the same work would have cost at pay-as-you-go rates, so read it as a measure of the work
• A token is a small chunk of text, roughly a word or a piece of one. Everything you type and everything Claude writes back gets counted that way
• The conversation is a room you rent, and you pay rent every turn, because the model has no memory of its own and re-reads the whole room from the top each time
• The kept copy of that room is a cache, and the caching discount makes re-reading roughly a tenth of the price. Every turn resets the timer, so a phone call or a careful review costs you nothing
• On a subscription that timer is an hour. The five minute figure people repeat is real, and it belongs to an API key setup rather than yours
• Your project notes file is read once when the session starts, so anything you add loads on the next clear, compact or restart
• A longer conversation does not make Claude sharper. It makes the room bigger, and you pay rent on all of it, every turn

---

Lecture: How do I catch the spend before it spikes? (Reading the Meter)

Nothing warns you, so the habit is glancing on purpose, the way you glance at a fuel gauge long before the light comes on.

🎯 Core Concepts:

• The ring holds two numbers: how full this one conversation has become, and how much of your plan you have spent everywhere you are signed in
• Glance at usage before a big task and again after it, and the difference tells you what that piece of work cost
• Your own setup has a cost. Every extra ability switched on and every long notes file gets sent at the start of every session, then pays rent on every turn after that
• For heavy users a flat subscription can be an extraordinary deal, and the cheapest plan is best understood as a paid trial. Feeling its limits is a signal to move up, not a sign you did something wrong

🚀 Actionable Items:

• Pick the two cost rules that fit how you actually work, rather than adopting all of them at once

---

Lecture: How hard should Claude think? (Effort Levels)

How hard Claude thinks before it answers, and where that thinking is worth paying for.

🎯 Core Concepts:

• At low, Claude stays narrow and literal and takes your request at face value. At max it slows down and considers the edges, the cases you did not mention
• Low effort is a quick sketch. Max effort is an architectural blueprint, worth every bit of it when the thing you are deciding is costly to reverse
• The ladder runs low, medium, high, xhigh, max, and ultracode at the top. High sounds close to the top and actually sits in the middle
• On the models you are most likely using, high is where you already are without choosing anything
• If an answer feels oddly shallow, moving one level higher is worth trying, because this setting does not always behave the way you expect

---

Lecture: Which brain should I pick for this job? (Choosing a Model)

Reaching for the most powerful model on everything feels safer, and it is the wrong instinct.

🎯 Core Concepts:

• The deep model is for planning and hard reasoning, a faster one for carrying out a plan you have already settled, and the lightest for simple repetitive passes
• Cheaper is not always cheaper. A lighter model looping for five minutes and never quite getting there costs more than the strong one finishing in a single turn
• Switching models mid-session makes Claude re-read the room for the new one, so that discount gets rebuilt from scratch every time
• Helpers Claude runs on its own quietly inherit the expensive model unless you tell them otherwise, and setting a cheaper one for them is a high-value adjustment
• The tool and the model are two different things from the same company. The tool is the workshop you work in, the model is the brain you have loaded into it

---

Lecture: Why do long sessions drift, and how do I stop it? (Staying on Track)

This failure is quiet, and quiet is what makes it dangerous. The work does not break in any obvious way, it drifts.

🎯 Core Concepts:

• Drift compounds because a fresh session has no reliable record of the decisions that came before it, so it re-derives them and moves a little further each time
• No amount of better prompting fixes it, and a bigger context window does not either. Long context does not mean good context
• When Claude tells you it pushed the code or the tests are green, Claude is grading its own homework. You check from outside, every time, and a drifting model reports success just as confidently as a right one
• When you correct something and it comes back wrong again, stop replying in that thread. It is now reading its own failed attempts as if they were the truth
• Hand off around the point where the room is half full, while the thinking is still sharp
• Record why you made a decision, not only the rule, so that later you can tell when it no longer applies

🚀 Actionable Items:

• One mission per session, handed off cleanly, beats one endless session every time

---

Lecture: How do I build a setup that is actually mine? (Your Own Workflow)

A cockpit you have not set up for yourself is still someone else's.

🎯 Core Concepts:

• Set a default model, using the reasoning about matching the brain to the job
• Pick the permission mode you want to sit in when nothing unusual is going on
• Add the line that stops Claude writing over your keys file, not only reading it, because a read rule does not cover a fresh file written straight over the top
• Write down each choice with one sentence on why you made it. Anyone can copy a setup, and the reason is what lets you change it when your work changes

🚀 Actionable Items:

• Watch the comma when you add a line to a list. Your new last line does not take one, and the line above it now does

---

Lecture: You're driving now (The 5 Things You Keep)

The difference between a passenger and an operator.

🎯 Core Concepts:

• You can read what your first project cost, and compare one session against another before a number surprises you
• You decided on purpose what Claude can do on its own and what it has to stop and ask you about first
• You chose your own setup with your eyes open about the cost and the risk, instead of hoping the defaults were fine
• A clean session does not automatically inherit any of those decisions, which is why the durable parts need a home you can inspect and maintain

This resource is part of

Claude Code: From Beginner to Pro Agentic AI Engineer

by Arun Nagarathanam

Take your skills to the next level with

AI GEO 2026: Get Recommended by ChatGPT, Google, Perplexity

by Arun Nagarathanam