Styles, custom instructions and memory

0:00 Make Claude yours

Every time you type use British spelling or keep it short, you are paying a small tax. Personalisation removes that tax, but only if you put each preference in the right place. In this lecture you will learn the layers of personalisation in Claude, how to build a custom style that sounds like you, and how to manage memory, which changed a lot in twenty twenty six.

0:29 Why layers matter

Why does it matter where a preference lives? Because rules in the wrong place either leak or clash. Put a client specific disclosure rule in your global profile, and it appears in work for a different client. Put your spelling preference in five Project instructions, and one day they disagree. Think of it like the settings on your phone. Some are system wide, some belong to one app. Getting the layer right means your preferences apply exactly where they should, and nowhere else.

1:05 The layers

Think in layers. Profile preferences apply to every chat, so they hold who you are, your spelling and general format preferences. Styles control how responses sound. Project instructions hold rules for one client or product. The chat holds task details. And memory holds facts Claude learns about you over time. The golden rule is that each rule lives in exactly one layer. When the same rule lives in three places, they drift apart and start contradicting each other.

1:39 Profile preferences

Profile preferences should be short, stable and non confidential. Something like, I am a growth marketer at a skincare brand selling in the UK and UAE. British spelling. Start with a one line answer. Prefer tables for comparisons. Flag assumptions. Give two options for copy. Do not put client names or anything sensitive here, because it travels into every chat you ever have.

2:06 Custom styles

Styles are where Claude starts to sound like you. Alongside presets like concise and formal, you can create a custom style from samples of your writing. Pick three to five of your best pieces in the same genre, describe the voice in one sentence, such as warm, direct, short paragraphs, occasional dry humour, and test it on a brand new topic. If it copies your phrases too literally, adjust the description. Keep separate styles for genuinely different voices, like your founder voice and your brand voice.

2:43 Memory in 2026

Memory evolved fast this year. It reached free users in March twenty twenty six. In July it moved to individual categorised entries you can actually view. In August you gained the ability to edit topics, plus a sensitive topics setting that controls whether sensitive subjects are remembered. Projects keep their own separate memory. Incognito chats are not saved to history or memory. And on Team and Enterprise plans, admins decide whether memory is available at all.

3:16 Treat memory like a notebook

Treat memory like a colleague's notebook about you. Useful, but worth checking. Once a month, open it, correct anything wrong, delete stale entries like an old job title or a client you no longer work with, and decide your sensitive topics setting. For conversations you do not want remembered at all, use incognito. One caution. Incognito controls your history and memory. It does not change your organisation's legal or contractual obligations about data.

3:48 Simple example

Here is a simple example. Add two lines to your profile preferences. Use British spelling. Start every answer with a one line summary. Now ask three unrelated things, a recipe substitution, a short email to a supplier, and help with a spreadsheet formula. Every answer starts with a single summary line, and every one says colour and organise, not color and organize. You wrote the rule once and it applies everywhere. That is the power of choosing the right layer.

4:23 Worked example: three voices in Abu Dhabi

An independent consultant in Abu Dhabi writes for three audiences. Government clients need formal, bilingual reports. LinkedIn needs a personal, conversational voice. Her paid newsletter is analytical. She sets profile preferences for spelling and format, creates two custom styles from five samples each, and gives each government client its own Project with formal instructions and disclosure rules. Memory stays on for general work, and anything involving confidential negotiations happens in incognito. The payoff showed up in small moments. When she drafted a LinkedIn post, it sounded like her. When she opened a government client's Project, the formal rules applied automatically. And when a negotiation got sensitive, she moved to incognito without thinking twice. Her monthly memory review takes five minutes, and she has never been surprised by an old detail resurfacing in the wrong place.

5:22 Pitfalls

Four pitfalls. Stuffing profile preferences with task rules that then fight your Project instructions. Custom styles that copy your distinctive phrases word for word, so ask for the voice, not the sentences. Forgetting memory exists and being surprised when an old detail reappears in a client draft. And assuming incognito changes your legal retention obligations. It does not.

5:47 Try this now

Try this now. First, write your profile preferences in under eighty words, who you are, spelling, and how you like answers structured. Second, pick three pieces of writing you are proud of, in the same genre, and create a custom style from them if your plan allows. Test it on a brand new topic and adjust the description if it copies your phrases too closely. Third, open memory, read every entry, and delete anything outdated or sensitive. Decide your sensitive topics setting while you are there. It takes about fifteen minutes, and every future chat benefits.

6:29 Watch me do it, part 1

Let me set up the consultant's layers. First, profile preferences. I type, independent brand strategist in Abu Dhabi, British spelling, one line answer first, tables for comparisons, flag assumptions. Under eighty words, nothing confidential. Next, styles. I choose to create a custom style and paste three of my best LinkedIn posts as samples. I describe the voice in one line, warm, direct, short paragraphs, occasional dry humour, never salesy. I name it LinkedIn me. Then I test it on a topic I have never written about, how hotels should handle AI chatbots, and check that it sounds like me without copying my old sentences.

7:14 Watch me do it, part 2

Now memory. I open it and read every entry. One says I work with a hotel group whose engagement ended in June, so I delete it. One topic is labelled retail, but my work is mostly hospitality now, so I edit the topic. I review the sensitive topics setting and choose not to have health related details remembered. Finally, I start an incognito chat and ask about a confidential negotiation strategy. Afterwards I check my history, and the chat is not there. Four layers, set in about ten minutes, and I know exactly what each one does.

7:56 Recap and next step

Recap. Put each rule in exactly one layer. Keep profile preferences short and non confidential. Build custom styles from real samples and test them. Review memory monthly, use the sensitive topics setting, and use incognito for anything sensitive. Your next step is the four part hands on in the lesson. Write preferences, create a style, clean your memory, and try an incognito chat.