Lesson 0.2.20.2.2 The controls, and the job each one does

In everyday life. A new kitchen has a few dials: the oven, the stove and the timer. You do not need to know how each one works inside. You need to know which one does the job in front of you, and how to tell when the job is done.

What you will learn. What each control does for you, and when to reach for it. How to check that an answer came from your file.

Why it matters. The same chat box can give you a guess or a checked answer. The control you pick makes the difference.

Most controls sit in one of three places: the sidebar, the message box, or Settings, which open from your name. The screens on this page are from 17 October 2026, and the places change often.

Figure 0.2.4. Where the controls are on each free plan. 0.2.3 lists them by name.

Talking to the model

When do you continue a chat, and when do you start a new one?

Continue a chat when you want the conversation so far to count. Before every reply, the app puts the earlier messages and their files back into the window, as 0.1 showed. Start a new chat for a new job, so that old messages do not steer the answer.

A new chat is not quite empty. The app's standing instructions, the rules its maker sets for every conversation, still load. Memory notes load if memory is on, and a project's files load if the chat is in a project.

Try it. In a chat that already has a few messages, ask "Quote my first message in this chat, word for word." Then ask the same in a new chat. The first chat can answer. The new chat can quote only that same question, because nothing came before it.

When do you pick a stronger model?

When the question is hard, or the answer matters. The model picker chooses which model reads the window. A faster model is fine for everyday questions. A stronger one, or a setting that lets it work through the problem first, helps with harder ones. It takes longer and uses more of your limit.

Figure 0.2.5. The stronger choice on each free plan. On Claude, Sonnet is stronger than Haiku. On ChatGPT, Think works through the problem first.

Try it. Open the model menu. Note which models your plan offers, and the name of its setting for working before an answer. Both AI vendors' names for them are in 0.2.3.

Giving it your material

How do you make it answer from your file?

Attach the file, and ask your question in the same chat. The file goes into the window for that chat, so the model can read it. Then check the answer. Ask it to quote the line it used, and find that line in your file yourself.

Figure 0.2.6. The + button in the message box attaches a file, on both apps.

The check matters because a reply from what the model learned in training can look the same as a reply from your file. Only the line shows where the answer came from.

What does a project do for you?

It saves you from attaching the same files, and typing the same instructions, again and again. A project holds files and instructions, and they load into every chat inside it.¹ Use one for a job you come back to, such as a client or a weekly report.

Figure 0.2.7. Making a project. After you make it, you add its files and instructions.

Try it. Make one practice project. Add a file to it and one instruction, such as "Answer in a short table." Then open two chats inside it. Both chats use the file and follow the instruction, without your attaching or typing them again.

What does memory do for you?

It carries notes about you from old chats into new ones, so you do not repeat yourself. You can read what it saved, and delete any note.

Figure 0.2.8. The memory switch on each app. Below it, you read what memory saved.

Try it. Open your memory settings, and read what the app has saved about you. On a new account it may still be empty. Delete any note you do not want.

Getting work done

When do you let it run code?

For anything you would check with a calculator: a total, a count or a table of numbers. With code execution, the model writes a small program and runs it, so the number is calculated instead of predicted, as 0.1.3 showed.

Look for your own numbers in the code it shows you. Some apps fold the code away, behind a label such as Analyzed or a small </> icon: click it. If one of your numbers is missing or wrong, so is the total. The result is evidence from outside the model, not proof. The code can still add up the wrong column.

Figure 0.2.9. The six amounts, in the code behind each total. On ChatGPT, the code opens from the </> icon.

How do you keep what it makes?

A result can land in three places: in the chat, as an Artifact the app keeps, such as a document or slides, or as a file you download. For anything you will use again, ask for the result as a document. The app keeps it as an Artifact, an output kept on its own, such as a document, a table or a small tool.² You can edit it, share it and find it again. Some apps make a file to download instead, such as a Word document. Your chat list keeps each chat with its files, too.

Figure 0.2.10. A document on each app. Claude opened one to edit beside the chat. ChatGPT made a Word file to download.

Try it. Ask for a one-page note about Brightline's approval rule, as a document. Change one sentence in it yourself, then find where the app keeps it.

Deleting a chat takes it off your chat list. The AI vendor deletes its own copy on its own schedule, as 0.2.3 shows. Chapter 1 shows which work survives a chat, and how to save what you need.

How deep should it look?

Pick the depth by the question. For one fact, such as whether Monday is a holiday, a quick web search in the chat is enough. For a question that needs many sources, research mode runs many searches and writes a report with its sources, where your plan has it (0.2.3 shows which). It takes longer, and you still check the sources. For a question about your own files, attach them, or reach them with a connector, such as Gmail or Google Drive, so the answer comes from your material and not from the web. Skills and connectors, from 0.1.3, come back in Chapters 2 and 8.

What do you check before you hand over work?

Four things. Are you signed in to the right account and workspace? Can it reach the tools and files it needs, and no more? Where will the result go? Which of its actions will ask you first?

Your first hour

Here is the first hour, with one invented file: Brightline's AP policy, version 3. Dave, the controller, owns it.

• 9:00. You open a free account, and make the two choices from 0.2.1: training and memory.
• 9:05. In a new chat, you ask whether a $7,800 invoice needs the controller's approval. The answer is a guess, because the policy is not in the chat.
• 9:10. You attach the policy, and ask again in the same chat. The answer quotes line 4.1: "Invoices over $5,000.00 need the controller's approval before payment." You find the line in the file.
• 9:15. You ask for this week's six invoices to be totaled by running code. It shows the code and the total.
• 9:20. You ask for the answer as a document, and find it again where the app keeps it.
• 9:30. You make a project for Brightline with the policy in it. Every new chat there starts with the policy.