How to Configure Claude for Yourself: Profile, Memory, and Projects

Practical materials for configuring how you work with models: Ailoot documentation.

Why does Claude start from scratch every time?#

A language model doesn't store you inside itself. It reads what it was given at the beginning of the conversation and works with that; to it, everything else simply doesn't exist.

So configuring Claude for yourself means putting the right text in one place once, so it gets pulled in automatically. You don't need to teach the model anything, and there are exactly four such places, scattered across different settings screens.

A lot of frustration with AI comes from this. People expect an assistant to remember them like a colleague. Instead, every morning brings a new intern, and the difference between “Claude is useless” and “Claude saves me a day a week” often comes down to twenty minutes of setup that nobody did.

Where exactly do the settings live: four different places#

Here's what it looks like in one table; the rest of this discussion will be about these four rows.

Anthropic's help page describes this section as follows.

The same page says that project instructions apply only to chats inside the project. Put information about you across the board in your profile, and everything about a specific piece of work in the project.

One caveat about the interface. Anthropic's personalization section is currently being revamped: in the same help article, the introduction talks about “styles,” while the middle has already been rewritten around skills. The article about styles has been removed from the collection of settings articles altogether. So if styles are still in your menu but your friend no longer has them, you just got a different wave of the update.

Profile settings are your personal layer: describe once how you think and what annoys you, and Claude stops suggesting the obvious. This layer alone already saves time.

How do you get Claude to interview you about yourself?#

Writing about yourself from scratch is hard. You open an empty field, and all you can think of is “entrepreneur, value clarity”—exactly the kind of useless string of words that changes nothing in the answers.

An interview solves this because answering is easier than composing. On its page for newcomers, Anthropic itself suggests asking Claude to interview you, in its list of “what to try next.” The ready-to-use prompt isn't there; it's below.

Interview me to gather text for the “Instructions for Claude” field in my settings.

Interview rules:

Ask exactly one question at a time and wait for my answer.

If my answer is vague, press for specifics. “I work in marketing” is a bad answer; ask: in what niche, how many people are on the team, who is the client, what do I do hands-on every day?

Don't praise my answers. No “great answer” or “got it, awesome.”

Don't ask again about something I've already told you.

Don't polish my wording. If I put it bluntly, write it that way.

What you need to find out:

What I do and how I earn money. Be specific. A job title alone isn't enough.

What tasks I bring to you most often.

What do I explain to you again and again from one conversation to the next? Is there anything I have to tell you twice?

How I prefer to receive answers: length, level of readiness, whether I want options or one answer.

What annoys me about AI responses. Get me to give specific examples; generalities aren't enough here.

What topics and phrasings to avoid.

When you're done, assemble everything into ready-to-use text of up to 400 words, in my voice, using simple phrases. Start with what I do, then my work tasks, then my response preferences, and finish with a list of prohibitions. No generalities like “values efficiency”—only what I actually said.

Start with the first question.

The logic is simple. A rule deserves a place in the settings only if you've already said it twice. Everything else is guesswork about yourself, and it takes up space without doing anything.

Next, put the resulting text where it belongs:

1. Click your initials in the lower-left corner and choose Settings.
2. Find the “Instructions for Claude” section.
3. Paste the text and save.
4. Check in a new chat that it was picked up (the section on checking below explains exactly how).

The account name is changed separately, in Settings → General, in the “Full name” field. These are different things: the interface sees the name, while the model reads the instructions.

What should you put in the instructions, and what should you leave out?#

Anthropic doesn't publish an official character limit. The constraint here is practical: the longer the text, the less reliably it is followed. For organizations, a limit is explicitly specified—3,000 characters for the entire company's instructions—so use that as a guide to the scale.

What works:

1. How you earn a living. A job title alone isn't enough. “I run a family law practice, have three assistants, and get clients through referrals” works; “lawyer” doesn't.
2. What you bring most often. Texts, calculations, document reviews, code. Claude will stop suggesting things you don't use.
3. Response format. Do you need ready-to-use text right away or a plan first, one option or three, with or without reasoning?
4. Specific prohibitions. The model will follow “Don't start answers with the words ‘great question.’” “Be concise” won't change anything.

What you can safely cut:

• Polite padding like “you are a world-class expert.” New models no longer respond to formulas like these, but they still take up space in every request.
• Situational rules for a single task. They belong in a project; in settings that apply everywhere, they get in the way.
• Contradictions. The Claude Code documentation says it plainly: if two rules contradict each other, the model may choose either at random. Read through your settings in full from time to time, because simply adding more at the bottom won't get rid of a contradiction.

Check each line this way: remove it and see whether the answer changes. If it doesn't, that line was decoration.

Here's what a completed text looks like. The example is composite and the niche is invented, but its structure is exactly what you should paste in.

I run a custom furniture studio, with six people in production and sales through word of mouth. The sales cycle is three weeks.
I handle measurements, estimates, and client correspondence; I outsource design.

What I bring most often:
- estimates and material cost calculations
- replies to clients objecting to price and timelines
- product listing copy
- reviews of supplier contracts

How to respond:
- give me ready-to-send text for the client right away
- keep it to no more than five paragraphs unless I ask for details
- round figures to the nearest hundred
- if information is missing, ask one clarifying question and wait for my answer

What not to do:
- don't start with “great question” or “of course”
- don't suggest hiring a marketer; I don't have the budget for it
- don't write “in light of the foregoing” or “it bears mentioning”
- don't give three options when I asked for one

Almost nobody fills in the last section, though it matters more than the other three. Prohibitions are the hardest things to formulate because you remember being annoyed, but not the exact phrase. That's why the interview prompt has a separate question for this, one that gets you to give examples.

How do you turn on memory, and what does it remember?#

Memory changed on July 10, 2026; it's in the release notes. Before, Claude summarized conversations into one overview once a day. Now, memory is a set of separate entries by category, which it reads and edits as the conversation goes along.

What goes in: your role and projects, your usual communication style, technical preferences, and details of current work. The help page separately notes that memory is geared toward work context, so personal details unrelated to work may not be saved.

Do three things right after turning it on.

1. Check the list of entries

   Settings → Memory shows everything Claude has recorded about you, organized by category. There's almost always a line or two in there you weren't expecting.

2. Clean up what's unnecessary

   You can edit it in two ways: through the “Tell Claude what to change or remove” field in settings, or directly in a conversation by telling Claude what to remove. The “Delete” button deletes the entry.

3. Make a habit of checking once a month

   Memory accumulates quietly. If you don't review it, six months from now there'll be a layer of junk from tasks that were completed long ago.

A separate note about private conversations. The ghost icon in the upper-right corner of a new chat turns on incognito mode: that conversation goes neither into memory nor history. It's available on all plans, including Free, but the icon isn't available inside projects, so incognito mode doesn't work there.

Why does memory sometimes make answers worse?#

The mechanics make sense once you know how memory works. Here's how: just before the request is sent, memory is added to it as ordinary text. It doesn't get inside the model or teach it anything.

So every unnecessary entry takes space away from the request itself and pulls the answer in its direction. Five precise lines work better than fifty random ones, and this also explains why reviewing matters more than adding.

Here's the order:

1. First, write your profile by hand, using the interview prompt. It's your text, so you know what's in it.
2. Turn on memory as the next step, as a supplement.
3. Once a month, open the list of entries and clean it out.

Then you'll have one layer you're responsible for and another that accumulates on its own. If the answers go off track, look at the second one first.

Memory reset, by the way, is irreversible. The help page is explicit: it deletes all entries, including project memory, and can't be undone. If you're unsure, start with a pause. Existing entries will be saved, but new ones will stop being created.

How are projects different from memory?#

Projects are available on all plans; on the Free plan, you can create up to five. A project contains three things: instructions, files, and chats. Instructions apply only to conversations inside that project; they don't carry over outside it.

Here's the breakdown:

For example, the owner of a coffee shop chain keeps one project for finance, another for hiring, and a third for social media. Each has its own instructions and files, while the fact that they can't stand answers longer than a screenful goes in the profile and applies everywhere.

When a project has more files than fit in the window, the extended project knowledge base kicks in. According to Anthropic's help page, it increases capacity by about ten times, but is available only on paid plans.

In addition to rules, project instructions should include four things:

1. The role you want Claude to take in this work. “You're the CFO of a small chain; you calculate conservatively and always show your work” sets the tone for every conversation in the project.
2. Stable facts you'd otherwise dictate again. Rental rates, team composition, location names, region. Anything that changes less often than once a quarter.
3. The output format for this specific work. A financial project needs a table; a writing project needs finished copy without padding.
4. Boundaries. What not to do in this project. For example, don't suggest tools you've already rejected.

Files belong there too, and they work better than any paraphrase. A price list, a contract, last year's report—Claude will answer based on them. Without files, it relies on general knowledge about your industry, and you'll notice the difference.

How can you check whether the setup worked at all?#

People usually check the setup incorrectly: they ask, “What do you know about me?” The model can piece together an answer to that question from anything, including the current conversation, so it proves nothing.

1. Open a new, empty chat

   Not a continuation of an old conversation, but a new one. Context has already accumulated in the old one, which will muddy the picture.

2. Ask a work-related question without explanation

   One that used to need an introduction. “Draft a reply to a client asking for a 20% discount”—and see whether you have to add what kind of business you run.

3. Check your prohibitions

   If you asked it not to start with “great question,” check the first line of the answer. It takes a second.

4. If it didn't work, check where you put it

   The setting may have gone into a project, while you're asking in an ordinary chat. Profile instructions apply everywhere; project instructions apply only inside the project.

In Claude Code, the check is shorter: /context shows a “Memory files” section listing what actually loaded into the session. If your file isn't there, further edits are pointless; first you need to figure out where it belongs.

Can you transfer memory from another AI service?#

Here's the process:

1. Ask your current service to list everything it remembers about you, verbatim and in one block.
2. Go to Settings → Memory and click “Start import.”
3. Paste the exported text.
4. Click “Add to memory”—Claude will break the text into separate entries itself.
5. Open the list of entries and review it. Some lines will almost certainly be junk from old conversations.

Two warnings. First, Anthropic itself labels the import experimental and warns that some entries may not transfer. Second, the import documentation talks about “other AI services” in general; it doesn't name specific services, so you'll have to review what was transferred by hand.

You can export it back out too: Settings → Capabilities, then the “View and edit your memory” link; or simply ask Claude to list the entries about you—it can do that.

How do you configure Claude Code for yourself?#

Claude Code knows nothing about your web settings. It's separate.

So you'll have to describe your personal preferences twice. Use the same text, though: put the interview prompt's result both in the settings field on the web and in a file on disk.

Files are read from general to specific and combined. None cancels out the previous one:

1. ~/.claude/CLAUDE.md—your personal preferences, applying to every project on this machine.
2. ./CLAUDE.md or ./.claude/CLAUDE.md—project rules, included in the repository and used by the whole team.
3. ./CLAUDE.local.md—personal to a specific project, placed in .gitignore.

If you're just getting started, run /init in the project folder. Claude will read the code and write a starter file with build commands and discovered conventions. If a file already exists, it will stay in place and Claude will suggest edits to it.

A detailed breakdown of what to put inside is in the separate article about configuring CLAUDE.md. The difference between the file and memory is covered in the article about Memory and CLAUDE.md.

Auto-memory works alongside these files. It's on by default and works like this: Claude writes its findings to the ~/.claude/projects/<project>/memory/ folder.

According to the Claude Code documentation, at the start of a session, the first 200 lines of MEMORY.md or the first 25 kilobytes are loaded, whichever limit is reached first. Anything below that isn't loaded at all, and a toggle in /memory turns off the whole mechanism.

There's a catch. Auto-memory lives on a specific machine and doesn't move between computers: if you work on a laptop and a desktop computer, they have two different memories. Syncing the project folder through the cloud doesn't help, because memory is stored in the home folder, which the cloud doesn't touch.

Half of the “it ignores me” cases can be diagnosed with /context right here. Usually it turns out that the file is in a neighboring folder or doesn't exist at all, while the person has been editing it for the third time in a row and expecting different behavior.

Why do long instructions stop working?#

The reason is simple, and it explains almost all the disappointment people feel with configuration.

Short, precise text works better than long, detailed text. If you want to add a fiftieth rule, it's more honest to remove three old ones first.

Second. If something must happen every time, without exception, text can't guarantee that in principle, because an instruction remains a request, however emphatic.

Claude Code has hooks for such cases: they run like ordinary commands at fixed moments and don't depend on what the model decides. The rule “run tests before saving” in text works some of the time; the same prohibition enforced through a hook works every time.

Third. Settings age along with models, and wording that changed behavior a year ago can sometimes get in the way in newer versions.

Where should rules that aren't always needed go?#

Take a contract review process: what to look for, in what order, what to write down—a page and a half of text.

You can't put that in your global settings; they'll bloat and stop working. Keeping it in your head and pasting it in by hand every time isn't much fun either.

A separate skill solves exactly this problem. Describe the process once and put it there; it loads only when the conversation turns to contracts. The other ninety conversations that month won't know about it or spend context space on it.

About styles. Responses used to be configured through a separate feature with that name. Anthropic is now consolidating it into skills: the styles section has been removed from the help collection, and skills have taken its place in the personalization article. The update is reaching users in waves, so some readers still have the styles menu.

Official documentation doesn't describe a mechanism for “upload your texts and Claude will learn your style,” either there or elsewhere. Describe the style you want in words, and add examples of your writing to the project as files.

What does Anthropic do with this data?#

Anthropic puts it this way.

If you put work information in your settings:

• These rules apply on Free, Pro, and Max, including when you use Claude Code with those accounts. They don't apply to enterprise plans or API use.
• If you delete a conversation, it won't be used for training.
• Incognito chats aren't used for training at all. They're retained for 30 days.

Essentially, you're keeping a file about yourself that's read every time you reach out. Treat it like a work document.

One more detail for the meticulous. The help article about how memory works says nothing about model training. That topic is covered on another page, so if you read only about memory, you won't learn about the five-year period.

What should you check after three months?#

Review:

1. Open your profile instructions and read them from top to bottom. Usually a third can be removed without losing anything.
2. Go through your memory entries. Delete everything related to completed tasks.
3. Try getting by without some of the settings. Temporarily remove what you've accumulated and see how the model does without it. The model has gotten smarter, while the workarounds remain and now get in the way.
4. Check that the files are actually loading. In Claude Code, use /context.

There's also a monthly report: Settings → Reflect shows what topics took your time, which day was busiest, and which hour you work most often. Launched on July 9, 2026, it works on Free, Pro, and Max when memory is on. It isn't available on enterprise plans; incognito conversations and health topics aren't included in the report.

Use it as a mirror. If the month's top topics don't match what's written in your settings, your settings describe the person you were in the spring.

Three months is simply the point at which the difference is usually noticeable. If you change jobs, someone joins or leaves the team, or your main stream of tasks changes, reread them right away rather than waiting for the calendar.

What else do people ask?#

Practical materials: Ailoot documentation.

More materials and tools: Ailoot.