Chapter 12: ChatGPT From Beginner to Advanced User
Read this first
Everything in this chapter with a product name or a number attached is a snapshot taken in the third quarter of 2026. ChatGPT’s pricing and model lineup changed repeatedly during the preceding year, and that pace is likely to continue. Some names and details may already have changed.
What will still be true is the shape: the capability categories, what each is for, and how to decide what you need. Read the structure, verify the specifics, and use Chapter 17’s method when the details have moved.
The three levels
Most people plateau at level one and conclude the product is a chatbot. It has not been that for some time.
Level one: conversation. Typing questions and getting answers. Where most users stay.
Level two: configuration. Making the tool persistently yours: instructions, memory, workspaces, files. Perhaps twenty minutes of setup, and it changes every conversation afterward.
Level three: delegation. Handing over multi-step work: research, agentic tasks, recurring automation. This is where the paid tiers earn their price, and where Chapter 9’s verification discipline becomes non-negotiable.
Level two: Configuration
Custom instructions
A standing note applied to every conversation. You filled it in during your first hour; here is what actually belongs there.
• How you want to be addressed and at what level of expertise
• Format preferences: default to prose, not bullets; no preamble
• Standing prohibitions: words, phrases, or habits you dislike
• An uncertainty instruction, when you’re not confident, say so explicitly rather than hedging
Keep it short. Long instructions get partially applied.
Memory
The tool retains details across conversations: your work, your preferences, ongoing projects. Genuinely convenient and genuinely a privacy surface.
Three things to know: how to see what it has stored, how to delete individual items, and how to turn it off. Check all three now. Memory capabilities have been expanding steadily, which means the amount of accumulated detail about you grows whether or not you think about it.
Projects
Grouped conversations sharing instructions, uploaded files, and their own context. Project conversations carry context across chats within the project, and project memory stays separated from your main memories in both directions.
If you have recurring work (a client, a course, a book, or a research area), set up a project once, load the background documents, write project-level instructions, and stop re-explaining yourself in every session.
Files, images, and voice
Upload documents, spreadsheets, and images; ask specific questions about them. The later file-work chapter explains where this fails on long documents and numerical data.
Voice conversation is worth trying even if it sounds gimmicky. For many people it becomes the primary mode for thinking out loud.
Level three: Delegation
Deep research
A mode that browses many sources over several minutes and returns a structured report with citations. It is a real capability and it is the single most misunderstood feature in the product.
Here is the necessary caution: it synthesizes from web content but does not independently verify facts — the report carries citations, and you still have to check the claims against the originals. A twenty-page report with sixty footnotes produces a powerful impression of thoroughness. The impression is not evidence. Citations can point to pages that do not support the claim attached to them.
Use deep research as an accelerator: it can find the territory and likely sources far faster than you would. Then read the sources that matter. The volume of output makes careful scrutiny of numbers, dates, names, quotations, and citations more important, not less.
Agent mode
The tool carries out multi-step tasks with reduced supervision: browsing, filling forms, assembling outputs, working across sites.
Agents sit at the current frontier and require unusual caution. An error in a single answer is visible and contained. An error at step three of a twelve-step autonomous task propagates silently through the remaining nine. Later guidance on agents includes what never to delegate.
Custom GPTs
Configured versions of the tool with fixed instructions, uploaded reference material, and a defined purpose, saved for reuse and optionally shared.
Worth building when you do the same specialized task repeatedly and have reference material that should always be in scope. Not worth building for occasional tasks: a good saved prompt does the same job with less overhead.
Scheduled tasks and recurring work
Operations that run at set times without you present: a recurring summary, a weekly search, a scheduled report. Useful, and worth reviewing periodically, since a task set up in March and forgotten is still running in August.
Coding and technical work
A dedicated coding agent capable of reading a codebase, proposing changes, running tests, and submitting them. If you write software, this is likely the most consequential thing in the product. If you do not, ignore it, but note that its capabilities keep extending into general document and data work, so the boundary moves.
Where ChatGPT shines
Held loosely, because these things converge and diverge with every release:
• One of the broadest feature sets among major consumer AI products, and frequently among the earliest major products to introduce new capability types
• The largest ecosystem: third-party integrations, shared configurations, published workflows
• Strong general-purpose performance across almost everything
• The most mature research and agentic tooling at the consumer level
Chapters 13 to 16 handle the comparison with Claude, Gemini, and Copilot properly. The short version: for many new users, ChatGPT is a strong default, though it is not the best at everything.
The subscription decision
Start free. Use it for two weeks. Then ask three questions.
Am I hitting limits? If you regularly run out of messages or get downgraded to a lesser model mid-task, that alone justifies the entry paid tier.
Do I need the specific paid features? Deep research, agent mode, higher file limits, the better model on demand. If you cannot name which one you want, you do not need it yet.
What is it worth in time? For roughly the price of a casual lunch, a paid AI subscription can provide a month of advanced capabilities. If it saves you even an hour each month, the arithmetic is usually favorable.
Lower-cost intermediate tiers now exist and are worth checking before committing to the standard one. The very expensive tiers are for people running agentic work all day; if you are wondering whether you need one, you do not.
Review it quarterly. Subscriptions accumulate silently. The Bonus Resource Library has a subscription-review worksheet, and an annual review keeps the decision current.
Privacy settings to check today
Three specific things, worth ten minutes:
Training. Consumer plans commonly default to allowing your conversations to improve the model, with an opt-out available. Business and enterprise plans generally do not train on your data by default. If work material will ever pass through your personal account, change this now.
Memory. Review what has been stored. Delete what should not be there.
Advertising. Free tiers have begun carrying sponsored content. Know whether you are seeing it, because a recommendation that is partly commercial is a different object from a recommendation that is not.
These three checks are worth completing before you finish the chapter. Chapter 25 builds the full framework.
Exercise 12.1: Set up one project properly
Pick something recurring in your life. Create a project, upload three or four relevant background documents, and write project-level instructions covering context, standards, and what you always want.
Then run a normal request inside it and compare to the same request outside it.
Exercise 12.2: Audit a deep research report
Run a deep research task on something you know well. When the report arrives, pick five citations at random and open every one.
Check that the source exists, that it says what the report claims, and that it actually supports the point. Record your hit rate. That number is your calibration for every research report you read from now on.
Exercise 12.3: Verify this chapter
Go to the official pricing and release notes pages. Find three things in this chapter that have changed since I wrote it.
Write them in the margin with today’s date. That is the beginning of a repeatable way to stay current.
Learn More About Artificial Intelligence
Learn more about the CLEAR, TRUST and LEARN methods for Artificial Intelligence