AI Guides › Step-by-step guides
By Nigel Guy · 7 min read
Most people spend their first week with Claude poking at it: a recipe, a poem, a tricky email, a "which is better, you or the other one?" test. It feels like learning, and it is, but every one of those chats starts from zero and teaches Claude nothing it keeps. Meanwhile three settings that would make every later week better sit untouched, because nothing forces you to open them.
The rule: spend week one on the three settings that carry over to every chat (memory, a written brief about how you work and sound, and dictation), and treat the testing as a by-product.
You need a Claude account and about an hour across the week, not in one sitting. The settings below are on the free plan except where flagged.
Cost, at time of writing: Free is £0. Pro is listed at $20 a month, or $17 a month billed annually (about £15 and about £13 at time of writing; check the £ price at checkout, as VAT and exchange rates move it). You do not need to pay for anything in this guide. The one paid-only item is "Search and reference chats", flagged in Step 1.
One caveat applies throughout: Anthropic renames and moves settings fairly often. The names below match the official help centre as checked on 2026-10-04. If a label differs on your screen, look for the closest match rather than assuming the feature has gone.
Memory is on by default for Free, Pro and Max accounts, so "turn it on" really means "check it, then tune it". On Team and Enterprise plans an owner controls whether it exists, so you may need to ask them.
Two controls to know about before you need them:
| Situation | What to do |
|---|---|
| One chat should leave no trace | Start a new chat, click + in the message box, switch Memory off, then send your first message. Or use the ghost icon for an incognito chat, which is saved to neither history nor memory. |
| You want memory to stop growing | Switch off Generate memory from chats and choose Pause memory (keeps what exists, makes no new entries). Reset memory deletes everything permanently and cannot be undone. |
Projects keep their own separate memory space, so work for one client or topic does not leak into another. Use that: one project per area of work.
Memory is Claude's guess about you, built from chats. A brief is what you tell it on purpose. The two do different jobs, so do both.
In the current help centre, account-wide guidance lives under your initials (bottom left), then Settings, then Instructions for Claude. Older guides call this "profile preferences", and an earlier "styles" feature has been folded into the wider personalisation tools, so do not be surprised if screenshots elsewhere look different. Per-topic guidance belongs in a project's own instructions (paid plans; free accounts are limited to five projects).
Do not write the brief from a blank page. Paste two or three real things you wrote, ones you were pleased with, into a chat and use this:
You are an editor who writes style briefs for AI assistants.
Context: I am [YOUR_ROLE] working in [YOUR_FIELD]. My readers are [YOUR_READERS]. I write in [BRITISH_OR_AMERICAN] English.
Below are [NUMBER] samples of my own writing that I am happy with:
[PASTE_SAMPLES]
Task: Write a brief of no more than 200 words that I can paste into Claude's account-wide instructions. It must cover: how formal I am, typical sentence length, words and phrases I use, words and phrases I avoid, how I open and close messages, and my formatting habits (lists, headings, bold).
Rules:
- Describe only what the samples show. Do not invent habits.
- Write instructions in the second person ("Write short paragraphs"), not descriptions of me.
- Quote at most three short example phrases from the samples.
- If the samples are too few or too similar to tell, say so and list what extra writing you need.
Before answering, check each instruction against the samples and delete any you cannot point to evidence for. Then give me the brief and nothing else.
Fill in: your role, field, readers, spelling variant, and the pasted samples.
Paste the result into Instructions for Claude, then test it with one real task and trim anything that made the answer worse. Keep it short: a long brief gets half-followed.
Dictation turns speech into text in the message box. You still read and edit before sending, and you get text back. That is the point of it: speaking is faster than typing for long, messy requests, and long, messy requests are the ones where Claude gets better context.
How to start it, per Anthropic's guidance:
Language: mobile dictation lists 11 languages, English and others, with non-English support described as beta. Change it under your initials, then Speech Input Language. Anthropic states mobile audio is deleted after conversion and not used to train its models; read the current privacy page if that matters for what you dictate.
Dictation is not voice mode. Voice mode is a spoken two-way conversation, better for thinking aloud or rehearsing; dictation is better when you want a written result.
A practice routine for the week: each day, dictate one real task as you would explain it to a colleague (the situation, what you have tried, what good looks like), edit the transcript for obvious mishearings, and send. Names and product terms are where transcription slips, so scan those first.