How to Use Claude, Step by Step

⚠️ Version note: This lesson describes a real interface, and interfaces change. Last checked: 24 July 2026. If a button isn't where we say it is, check the "If it moved" note under that step — the logic stays the same, only the location shifts.

What you'll have at the end

• An account, set to the right model
• One Project, configured around your actual work
• A five-part prompt template you'll reuse constantly
• A reflex for noticing when the output is wrong

This is not a "ask Claude a question" lesson. Most people use it like a search engine and stop there. Everything that matters starts after that point.

Prerequisite: AI-01. Learning the interface without knowing what the model is doing is like learning the steering wheel and not the engine.

1. Go to claude.ai and create an account.
2. On the left you'll find three things: your conversation list, Projects, and search.
3. The conversation area sits in the middle, with the message box at the bottom.
4. Next to the message box: attach file (paperclip) and the model selector.

✅ Check: Click the model selector. If you see more than one model name, you're in the right place.

If it moved: The model selector is sometimes under the message box, sometimes in the top bar. Look for the words "Opus", "Sonnet" or "Haiku".

Most people skip this step and then conclude that "AI is dumb." Models exist for different jobs.

Practical rule: default to the middle tier. If an answer comes back shallow, switch to the strongest model and ask the exact same prompt again. See the difference yourself — until you've run that comparison once, you don't actually know what the models do differently.

✅ Check: You've asked one question to two different models and noticed a difference.

Here's what roughly 90% of people write:

❌ "tell me about presentations"

Here's the structure that changes the output. Five parts:

[1 ROLE]        You're a presentation coach with 10 years of experience.
[2 CONTEXT]     I'm in year 11. I have to give an 8-minute biology presentation
                to a class of 40. Topic: antibiotic resistance. The class isn't
                interested in the topic; the teacher grades on depth.
[3 TASK]        Draft me a skeleton for the talk.
[4 FORMAT]      Slide by slide. For each slide: title, one-sentence core message,
                what's on screen, what I say. As a table.
[5 CONSTRAINT]  Maximum 8 slides. No slide gets more than 15 words of text.
                Don't suggest generic stock imagery.

Now run both and compare. This is the highest-return ten minutes in the lesson.

Why it works: the model is predicting the next token. Every constraint you add narrows the space it's predicting into. A vague prompt gets an average answer, because average is the most probable prediction.

✅ Check: You've compared a one-line prompt against a five-part prompt on the same topic, and the second answer is noticeably more usable.

Save it: paste that five-part template somewhere you'll find it. You'll fill it in for every serious task from now on.

The paperclip takes PDFs, Word docs, spreadsheets, CSVs, images and code files.

Most people upload something and type "summarize this." A summary is the lowest-value thing you can get out of a document. Try this instead:

I've uploaded a 40-page report.

1. What's the report's central claim, in one sentence?
2. Which piece of evidence supporting that claim is the weakest, and why?
3. What are three questions this report should answer but doesn't?
4. Write me 5 exam questions this report could produce. Don't give the answers.

The difference: a summary is passive, this is interrogation. The first one saves you from reading. The second one means you actually understood it.

Context window: there's a limit to how much text the model can hold at once. On very long files it can lose detail from the beginning. The symptom: answers go vague, or it claims something isn't in the document when it is. The fix: split the file, or anchor your question to a section — "looking only at section 3…"

✅ Check: You've uploaded a PDF and asked all four questions. The answers reference real content from the file.

In a normal conversation, every new chat starts from zero. You reintroduce yourself every time. Projects fix that: you give the context once, and every conversation inside that Project has it.

1. In the left menu: Projects → New project
2. Name it — e.g. "School — Year 11"
3. In Project instructions, write the context that never changes:

I'm in year 11, on the science track.
My goal: [your target university / course].
Weak areas: organic chemistry, reading comprehension.
When I ask for help studying: don't just give me the answer. Make me attempt
it first. Only solve it if I get stuck twice.
Keep answers short. No motivational filler.

4. In Project knowledge, upload the sources you keep coming back to: the syllabus, your notes, documents your teacher gave you.

Now every conversation you start inside this Project already knows all of it.

Set up at least three:

✅ Check: You started a new chat inside the Project, asked something without introducing yourself, and the answer came back like it already knew you.

If it moved: "Projects" may be called "Workspaces" or something similar. What you're looking for is a container that holds persistent instructions + persistent files.

Ask for long code, a document or a table and the output opens in its own panel (an Artifact). That means the output isn't chat text — it's a document you can work on.

The real skill here is iteration. Don't try to write the perfect prompt up front. Get a rough version, then narrow:

Round 1:  "Make me a budget tracker."
Round 2:  "Separate income rows, add a monthly total."
Round 3:  "Turn the cell red if a category goes over budget."
Round 4:  "Add a yearly summary sheet."

Four rounds gets you somewhere better than one enormous prompt would — and because you watched each change happen, you actually understand the thing you ended up with.

✅ Check: You've developed one output across at least three rounds.

A long conversation is a confused conversation. Three signs you need a new one:

1. The topic changed. Don't ask a coding question in the middle of studying chemistry — the model drags the earlier context along as noise.
2. The model is stuck on a mistake. If it repeats the same error a third time, stop correcting it and start a new chat. That wrong answer is now part of the context.
3. The conversation got long. It starts losing your original instructions.

Rule: one conversation, one task.

The model can produce confident, wrong information. This is called hallucination, and it isn't a bug — it's inherent to the technology. The model generates what's probable, not what's true.

Never use these three without checking:

Quick test: ask the model "which claim in that answer is the weakest, and why?" It critiquing its own output is more reliable than its first answer.

Full protocol: AI-07 — Hallucination Hunting.

Six common mistakes

1. Using it like a search engine. One-line question → average answer. Give context.
2. Not choosing a model. Hard job to the fast model, trivial job to the heavy one.
3. Never setting up a Project. Typing the same context 50 times by hand.
4. Accepting the first output. Calling it bad without a single round of iteration.
5. Using it unverified. Especially numbers and sources.
6. Outsourcing the thinking. Taking the result and handing it in without understanding it — that finishes your homework and finishes you.

Exercise (30 min, verifiable output)

Task: turn Claude into a study partner for a real exam or assignment you have coming up.

1. Create a Project called "School" and fill in the instruction template above with your details.
2. Upload your notes or source material to the Project.
3. Open a new chat and ask for this:

   Using the source I uploaded, build me a 10-question exam that increases in difficulty. Ask the questions one at a time and wait for my answer. When I get one wrong, don't tell me the answer — tell me which concept I missed and ask again. After all 10, tell me which topics I'm weak on and what I should study.

4. Take the whole exam.
5. Put the resulting weak-topic list into the spaced repetition schedule from LN-01.

✅ Finish check: you have (a) a working Project, (b) a completed exam session, (c) a personal list of weak topics.

Cheatsheet

PROMPT SKELETON
  Role        → who should it answer as
  Context     → situation, constraints, audience, level
  Task        → exactly what you want
  Format      → table / bullets / code / length
  Constraint  → what it must not do

PICKING A MODEL
  Hard reasoning, long code  → strongest model
  Everyday work              → mid model
  Fast and simple            → light model

START A NEW CHAT WHEN
  the topic changed · it repeated the same error 3× · the chat got long

NEVER USE UNVERIFIED
  numbers · dates · sources · law · health · tax

AFTER UPLOADING A FILE, DON'T ASK
  ❌ "summarize this"
  ✅ "central claim / weakest evidence / what's missing / write exam questions"

Next lesson: AI-04 — The Anatomy of a Prompt (L1) Related: AI-03 How to Use ChatGPT · AI-07 Hallucination Hunting · AI-10 Build Your Own Assistant Path: The Rehearsal Room — 1/6 · AI Operator — 2/6