How to Use ChatGPT, Step by Step

⚠️ Version note: this describes a real interface and interfaces change. Last checked 24 July 2026. If a control isn't where this says, check the "If it moved" note under that step — the capability persists even when the menu doesn't.

What you'll have at the end

• An account with custom instructions that stop you re-explaining yourself
• One Custom GPT built around something you do repeatedly
• A spreadsheet or CSV actually analysed, with a chart you didn't make by hand
• A clear sense of when to reach for this rather than Claude

Prerequisite: AI-01. Same reason as always — the interface makes sense once you know it's predicting from context.

On comparisons: this lesson doesn't argue ChatGPT is better or worse than Claude. Use both. They differ in ways that matter for specific jobs, and the last step covers where each one wins.

1. Go to chatgpt.com and sign in.
2. Left sidebar: conversation history, and above it a control for GPTs.
3. Bottom of the screen: the message box, with attach (paperclip) and the model selector.
4. Bottom-left: your account menu — this is where settings live, and you'll need it in step 2.

✅ Check: open the model selector. If you see more than one model listed, you're where you need to be.

If it moved: the model selector migrates between the top bar and just above the message box. Look for model names rather than a fixed location.

Almost nobody sets these, and they change every conversation you will ever have.

1. Account menu → Settings → Personalization → Custom instructions
2. Two boxes. Fill both.

Box 1 — what it should know about you:

I'm in my second year studying [subject] at [university].
I'm comfortable with [what you already know] and weak on [what you don't].
I write a lot of [essays / code / reports].
English is my [first / second] language.

Box 2 — how it should respond:

Be direct. No preamble, no "great question", no summarising my question
back to me.
Don't hedge every sentence. If something is genuinely uncertain, say so
once and move on.
When I'm learning something, don't hand me the answer — make me attempt
it first and only solve it if I'm stuck twice.
If you use a technical term I might not know, define it in one clause.
Keep answers as short as the question allows.

That second box is doing the work of a prompt you'd otherwise retype forever. The "don't hand me the answer" line is the one to keep — it converts the tool from something that does your work into something that makes you better at it, which is the difference between using AI and outsourcing to it.

✅ Check: start a brand-new chat and ask something you'd normally have to set up. The answer should arrive already shaped — shorter, no preamble, pitched at your level.

If it moved: this may sit under "Personalization", "Customize ChatGPT", or inside a profile menu. What you're looking for is persistent instructions applied to every conversation.

The selector isn't decoration. Models differ in what they're built to do, and the names change often — so learn the categories, not the labels.

The habit worth forming: when an answer comes back shallow, don't rephrase — switch to the stronger model and send the identical prompt again. Until you've run that comparison once, you're guessing about what the models actually differ on.

✅ Check: you've sent one non-trivial question to two different models and can name a difference.

This is where ChatGPT has a genuine edge worth knowing about: it can write and run code on a file you give it, then show you the result.

1. Get a real CSV or spreadsheet — your spending export, a dataset for your course, anything with a few hundred rows.
2. Attach it with the paperclip.
3. Ask for analysis rather than description:

This is [what the file is]. Before analysing anything:

1. Tell me what's actually in it — columns, row count, date range.
2. Flag data quality problems: missing values, duplicates, anything
   that looks like an entry error.
3. Then: what are the three most interesting patterns here, and how
   confident are you in each?
4. Chart the clearest one. Explain what the chart does and doesn't show.

Show me the code you run.

That last line matters. When it writes code, the code is the audit trail — you can see what it actually computed rather than trusting a number in a sentence. This is a genuine reduction in the AI-07 problem: a number that came out of code you can read is a different kind of claim from a number it recalled.

✅ Check: you have a chart generated from your own file, and you can point at the line of code that produced the figure underneath it.

If it moved: this capability appears under names like "Advanced Data Analysis", "Code Interpreter", or just happens automatically on upload. Look for it writing and running Python, not just describing your file.

A Custom GPT is a saved configuration: instructions plus files, kept and reusable. The equivalent of Claude's Projects, from AI-02.

1. Sidebar → GPTs → Create
2. There's a conversational builder and a Configure tab. Go to Configure — the builder is slower and vaguer than typing what you want.
3. Fill in:
   • Name — e.g. "Essay Critic"
   • Instructions — the real content:

You critique draft essays for a second-year [subject] student.

When I paste a draft, do this in order:
1. State my thesis back in one sentence. If you can't find one, say so
   first — that's the most important thing you can tell me.
2. For each paragraph: what claim is it making, and does the evidence
   support that claim?
3. Name the weakest paragraph and why.
4. List three questions an examiner would ask that I haven't answered.

Never rewrite my prose. Don't give me replacement sentences.
Don't praise. Skip anything that's working and spend the space on
what isn't.

• Knowledge — upload the marking criteria, module handbook, a past essay that scored well

4. Save it. It's now in your sidebar permanently.

Build three, for the three things you genuinely repeat. The instruction that makes these good is almost always a prohibition — "never rewrite my prose", "don't praise", "don't give me the answer". Without those you get a flattering assistant that does your work badly.

✅ Check: you've run a real draft through your GPT and got structural criticism rather than compliments.

If it moved: may be called GPTs, Custom GPTs, or sit under a "Explore/Create" entry. You're looking for saved instructions + uploaded files, reusable across conversations.

ChatGPT can carry facts between conversations. Useful, and worth managing rather than ignoring — because it quietly shapes answers you'll never trace back.

• Settings → Personalization → Memory shows what it has stored. Read the list. It's usually more than you'd guess.
• Delete anything wrong. A wrong stored fact propagates silently into future answers.
• Use a temporary chat for anything you don't want retained.

The failure mode: it remembers something from a one-off context — a project you've finished, an assumption you've moved past — and keeps applying it. When answers start feeling subtly off-target, check memory before blaming the model.

✅ Check: you've read your stored memories and deleted at least one thing that shouldn't be there.

Honest, and it will age — recheck it rather than trusting this table forever.

The real answer: use both. They cost about the same, they're strong in different places, and having two lets you do the single most useful check available — same prompt, two models, compare. Where they agree you're on firm ground; where they diverge is exactly where you should be verifying.

✅ Check: you've run one identical prompt through both and noticed where they differ.

Six common mistakes

1. Never setting custom instructions. Retyping context forever when two boxes would have fixed it.
2. Leaving the model on default. Sending hard reasoning to a fast model, then concluding it can't reason.
3. Asking for a summary of an uploaded file. Lowest-value output available — interrogate instead (AI-02, step 4).
4. Building a GPT with only positive instructions. No prohibitions means a flatterer.
5. Ignoring memory. Stale stored facts steering answers you can't account for.
6. Trusting numbers because code produced them. Code you didn't read is still a claim. Read it.

Exercise (40 min, verifiable output)

1. Write both custom instruction boxes, including the "don't hand me the answer" rule.
2. Take a question you asked last week and re-ask it in a fresh chat. Note what changed.
3. Upload a real data file. Run the step 4 prompt. Save the chart.
4. Build one Custom GPT for something you do repeatedly. Include at least two prohibitions.
5. Run something real through it and note whether it gave criticism or compliments.
6. Open memory, read it, delete one stale entry.
7. Send one identical prompt to ChatGPT and Claude. Write two sentences on where they diverged.

✅ Finish check: custom instructions live · a chart from your own file with the code visible · a working Custom GPT with prohibitions · a memory list you've pruned · one two-model comparison.

Cheatsheet

SET ONCE, PAYS FOREVER
  Custom instructions → who you are + how to answer
  Best line: "don't hand me the answer, make me attempt it first"

MODEL CHOICE
  reasoning model  → maths, logic, debugging, multi-step
  flagship         → most things
  fast/small       → rewrites, formatting, bulk

DATA FILES
  don't ask for a summary. ask for:
  what's in it · quality problems · 3 patterns + confidence · one chart
  ALWAYS: "show me the code you run"   ← the audit trail

CUSTOM GPT
  Configure tab, not the chat builder
  instructions + knowledge files
  the good ones are mostly PROHIBITIONS:
    never rewrite my prose · don't praise · don't give me the answer

MEMORY
  read it · delete what's wrong · temporary chat for one-offs

BEST AVAILABLE CHECK
  same prompt → two models → compare. Divergence = go verify.

Sources

1. OpenAI — ChatGPT documentation: custom instructions, GPTs, memory, data analysis
2. OpenAI — model capability and selection documentation

Next lesson: AI-04 — The Anatomy of a Prompt (L1) Related: AI-02 How to Use Claude · AI-05 Which Model for Which Job · AI-07 Hallucination Hunting · AI-10 Build Your Own Assistant Path: AI Operator — 3/6