Claude: Zero to Hero
Foundations · module 1 of 8

Your first real conversation

You are here · All paths · Free plan · 45–60 min · Assumes What Claude actually is.

What you’ll learn: how to drive a Claude conversation deliberately rather than hopefully — including the interface controls most people never touch.

If you only read one thing

Most people use Claude like a search box and get search-box results. The single change that fixes this: write the prompt you’d send a capable new colleague who knows nothing about your situation. What you want, why, who it’s for, and what “good” looks like. That’s forty seconds of typing and it replaces three rounds of “no, shorter.”

Two habits worth building immediately:

The interface, control by control

Open claude.ai or the desktop app. Here’s what each thing does and when it matters.

Find these before you read on. Open Claude in another window and locate each one — this takes two minutes and makes the rest of the module concrete rather than abstract. Positions shift as the app changes; if something isn’t where this says, look nearby rather than assuming it’s gone.

Control Where to look What it looks like
Message box Bottom centre The big text field you type into
Model picker In or just above the message box The current model’s name — “Sonnet”, “Opus” — as a small dropdown
Effort / thinking toggle Same dropdown as the model picker, or Settings A slider or list: low / medium / high
Attachment button Left end of the message box A paperclip or +
Search and Research The + or tools menu by the message box Toggles labelled “Web search” and “Research”
Edit a message Hover over a message you sent A pencil icon appears at its edge
Retry Hover under a message Claude sent A circular-arrow icon
Settings Your name or avatar, bottom-left Standard settings menu

If you can’t find one, the fastest route is to ask Claude itself: “Where is the model picker in the app I’m using right now?”

The message box

Type, hit Enter. Shift+Enter gives you a newline without sending — use it constantly, because multi-line prompts are almost always better than single-line ones.

The model picker

Sits near the top or beside the input. Switching mid-conversation is allowed; the transcript carries over.

Rules of thumb:

Covered properly in Models, effort, and thinking.

Effort / thinking controls

Recent models let you set an effort level. Higher effort means more internal reasoning before answering: better on hard problems, slower and more token-hungry on easy ones.

Default is fine for most things. Raise it for maths, architecture decisions, subtle debugging, and multi-constraint planning.

The attachment button

Files, images, screenshots. Drag-and-drop and paste both work. See Files, images, and vision.

Web search and Research

Two different things:

See Web search and Research.

Editing and branching

Hover any message you sent and you can edit it. This does not append a correction — it rewrites history and re-runs from that point, discarding everything after.

This is the most underused feature in the product. When a conversation goes wrong, don’t argue with Claude in follow-up messages — that pollutes the context with the wrong answer and your correction. Go back, fix the original prompt, re-run.

You can also retry a Claude response to get a different generation from the same prompt.

Projects

A container with its own knowledge base and standing instructions. Covered in Projects.

Concepts: how to actually run a conversation

Start over more than you think you should

A conversation is a context window that only grows. Once it contains a wrong answer, a tangent, and three corrections, every subsequent response is being generated against that mess.

Heuristic: if you’ve corrected Claude twice on the same point, or you’ve changed topic entirely, start a new chat. Paste in the two or three things worth keeping.

Front-load the context, back-load the question

For anything with substantial input material:

[long document / data / code]

---

Given the above: [your actual question]

Not the other way round. Anthropic’s testing shows queries at the end can improve quality by up to 30% on complex, multi-document inputs.

Ask for the shape you want

Claude will pick a format if you don’t. Naming one costs you eight words and removes an entire round trip.

“…as a table with columns Risk, Likelihood, Mitigation.” “…in three paragraphs of prose, no bullet points.” “…as a numbered checklist I can print.”

Let it ask you questions

For anything ambiguous, this is the single highest-leverage sentence you can add:

“Before you start, ask me up to five questions that would materially change your answer.”

You’ll be surprised how often it identifies the thing you forgot to mention.

The prompt that works is the one you’d send a competent new hire who has no context on your situation. That means: what you want, why, who it’s for, what “good” looks like, and any constraints.

An anatomy of a good first prompt

Bad:

write a project update

Good:

Write a project status update for my engineering team’s weekly email.

Context: we’re building a payments integration. We shipped the sandbox environment this week, hit a delay on PCI review (now expected 2 weeks late), and started work on the refund flow.

Audience: 12 engineers plus two product managers. They’re technical and busy.

Tone: direct, no corporate padding. Lead with the delay because that’s what affects their planning.

Length: under 200 words. Plain prose, no bullet points.

The second one takes forty seconds to write and saves three rounds of “no, shorter” and “no, less formal.”

Try it

Exercise 1 — Feel the difference. Send the bad prompt above. Read the output. Start a new chat and send the good one. Note specifically what changed: not just quality, but how much of your job Claude did for you.

Exercise 2 — Practise editing over arguing. Ask Claude for a 500-word blog intro on a topic you know. When it’s not right, resist replying. Instead, edit your original message to add the missing constraint and re-run. Do this three times. Notice the output gets cleaner each time, and the conversation stays one turn long.

Exercise 3 — The question-first pattern. Pick something genuinely underspecified from your real work (“help me plan the offsite”, “review my pricing page”). Send it with: “Before answering, ask me up to five questions that would materially change your response.” Answer them. Compare against what you’d have got without.

Worked example — if you're not sure what to use **Send this:** > Help me plan a team offsite. > > Before answering, ask me up to five questions that would materially change your response. **Claude comes back with something like:** > 1. How many people, and are they all in one location currently? > 2. What's the primary goal — strategy, relationship-building, or a specific decision that needs making? > 3. How long: half a day, a full day, or multiple days? > 4. What's the budget range, and does it need to cover travel? > 5. Has the team done one before, and if so what went badly? **What to notice.** Question 5 is the one you wouldn't have thought to include, and it's the one that changes the plan most. That's the entire value of this pattern: it surfaces the constraint you were carrying implicitly. Answer all five and the plan you get back will be usable rather than generic. Compare: send "help me plan a team offsite" cold, in a fresh chat, and you'll get a plausible agenda for a team that isn't yours.

Exercise 4 — Format control. Ask the same question five ways: as a table, as prose, as a checklist, as a decision tree, as a one-line answer. Save the phrasings that worked into your prompt journal.

Exercise 5 — Deliberately fill the window. Have a long, meandering conversation — 20+ turns, several topic changes. Then ask a precise question about the first topic. Then start fresh, paste only the relevant part, and ask the same question. This is your proof that context hygiene matters.

Checkpoint

Going deeper