The 10-minute Claude setup that actually knows you

You are using a workspace tool like a search box

Most people open Claude, ask one question, get one answer, close it, and open a brand new chat tomorrow. That is search-engine behaviour, and it is why the answers read like search results. Claude is built to work with accumulated context: who you are, what you are working on, how you sound, what you never want it to do. None of that arrives on its own. You put it in once, in a place that persists, and it applies from then on.

The other half of the problem is what Claude does that you never asked for: the disclaimers, the restating of your own question back to you, the tidy summary at the end of every answer. Those are default behaviours, and they are switchable. Step five is where most of the annoyance goes away.

The five steps, in order

1. 📁 Start a Project, not a chat. In the Claude sidebar, open Projects and create a new one. Name it after the thing it is for, like "my content" or "client work", not "general". A Project is a workspace that holds instructions and files, and every chat you start inside it inherits all of that. This is the container the other four steps live in, so do this one first even if you stop there for today.
2. 🎤 Tell Claude who you are, properly. Open a chat inside the Project and answer the interview prompt below. Do not write three bullet points. If typing a long answer feels like work, use dictation on the mobile app: tap the microphone icon on the right of the chat input and talk for two minutes (Anthropic's dictation guide). Talking gets you five times the detail typing does, and detail is the entire point of this step.
3. 📌 Turn that into instructions and save them. A great answer sitting in a chat is still trapped in that chat. Ask Claude to compress everything you just said into a tight brief, then paste it into the Project's instructions field so it loads at the start of every conversation in there. If you want it to apply everywhere instead of just this Project, put it in your profile instructions: click your initials in the lower left, open Settings, and use the "Instructions for Claude" field (personalization features). Most people want both, a short account-wide version and a detailed per-Project one.
4. ✍️ Feed it five real samples of your writing. Upload them to the Project's knowledge, or paste them into a chat. Real ones: a long email you actually sent, a document, a post that did well, a couple of long messages to a friend. Mix formal and casual, because the gap between them is where your actual voice lives. Then run the voice prompt below so Claude extracts the pattern instead of vaguely imitating it.
5. 🚫 Add behaviour rules, the step most people miss. This is the one that changes the daily experience most. Tell Claude what not to do: no disclaimers, no restating your question, no filler summary at the end, no asking permission before doing the obvious next step. Paste the starter block below into your Project instructions underneath your context brief, and add a line every time something irritates you.

The four blocks you paste

Start with the interview. Answer it out loud if you can, and answer it like you are explaining yourself to a smart new hire rather than filling in a form:

I'm setting you up so you actually know me and I never have to explain myself at the start of a chat again. Interview me before you write anything.

Ask me a few questions at a time, and keep going until you have a real picture rather than a job title: who I am and what I actually do day to day, what I'm working towards over the next six to twelve months, who my audience or customers are and what they care about, the tools and platforms I work in, the kinds of tasks I'll bring you most often, and how I like to be talked to. When an answer is thin or generic, push back and ask me for the specific version instead of accepting it. Do not summarise or start writing until you have run out of useful questions.

Then turn the interview into something permanent:

Now turn everything I've told you into a set of instructions I can paste into my project instructions field, so every future chat in here starts already knowing this.

Write it in second person, addressed to you, in the format "here is who I'm working with and how to help them". Keep it under 400 words, because long instructions get skimmed rather than followed. Lead with the facts that change your answers most, and cut anything that sounds nice but would not alter a single response. Where I gave you a preference, write it as a rule you follow rather than a fact you know. End with a short list of things you should ask me about rather than assume.

Next, the voice extraction. Attach or paste your five samples in the same message:

Here are five things I've genuinely written. Analyse how I write, then write in that voice from now on.

Go past "casual and friendly". I want the mechanics: my average sentence length and how much I vary it, whether I use contractions, the punctuation I lean on and the punctuation I never touch, how I open and how I sign off, the specific words and phrases that recur, my paragraph length, and whether I explain first or state the point first. Then list the things I clearly avoid. Give me the analysis as a short style guide I can save, and prove it works by rewriting [paste a paragraph of generic writing] in my voice so I can check whether you actually got it or just flattered me.

And the behaviour rules. This block is meant to be pasted straight into your Project instructions, underneath your context brief:

How to respond to me:
- No disclaimers, no "it depends on your situation", no hedging on things I did not ask you to hedge on.
- Do not restate my question back to me. Start with the answer.
- No summary paragraph at the end. When the answer is finished, stop.
- No praise openers. Skip "great question" and start working.
- When I ask for something, do the obvious next step too rather than asking permission for it.
- If I'm wrong about something, say so directly in the first line, with the reason. Do not soften it into a suggestion.
- If you are unsure or the information could be out of date, say which part and how confident you are, rather than writing confidently and hoping.
- Match the length of your answer to the size of the question. A one-line question gets a one-line answer.
- Write in my voice using the style guide in this project. Sentence case, no em dashes.

The two-minute test

Setups fail quietly, so verify rather than assume. Open a brand new chat inside the Project and ask for something you would normally have to set up first, for example "draft the email I owe [person] about [thing]". Three things should be true:

The habit that keeps it good

This setup decays if you never touch it again, because your work changes and the instructions do not. The maintenance is genuinely small, and it is one line at a time.

1. ➕ Add a rule the moment something annoys you. When Claude does the thing you keep correcting, open the Project instructions and write the rule right then, in one line. That is the whole loop. Ten of those over a month and it stops feeling like a tool you fight.
2. 🗂️ Split into a second Project when contexts collide. The moment you notice client context bleeding into personal work, or one Project's instructions contradicting what you need for another job, make a second Project. Each one has its own instructions and its own knowledge, which is exactly what keeps two very different kinds of work from muddying each other.
3. 🔄 Refresh the writing samples every few months. Voice drifts. When you notice your recent writing is sharper than what Claude imitates, swap in three newer samples and rerun the voice prompt. Two minutes, and it is the difference between sounding like you now and sounding like you last year.

Proof

Anthropic, what are Projects (plan availability and the five-project free limit)
Understanding Claude's personalization features (profile instructions, project instructions, styles, skills)
Use dictation on Claude mobile
Use voice mode