Back to Tutorials
ProductivityOperations

Where Claude's Instructions Live: Profile, Projects, Skills

6 min read


The first time you notice you have typed "keep it under 150 words and skip the intro" for the fifth time, you start looking for a place to write it once. Claude has three such places, and they are easy to confuse because each one is, in a sense, "where the instructions go." Putting a rule in the wrong one produces strange results: a client's tone leaking into your personal chats, or a habit you wanted everywhere that only works in one project.

This article maps the three, tells you what belongs in each, and shows how they combine. The Complete Beginner's Guide to Claude introduces them briefly. Here is the long version.

The three places

Anthropic's personalization page separates them by reach.

  • Profile instructions are account-wide. They apply to all of your conversations. You set them under your initials in the lower left, then Settings, then "Your Claude." Anthropic's suggested contents are preferred methods, common terms, typical scenarios and communication preferences.
  • Project instructions apply only to chats inside one project. They suit project-specific context, workflow guidelines, task requirements and roles for Claude.
  • Skills add specific behaviors or capabilities to conversations. They can adjust tone and format, apply patterns based on your own writing, or add specialized expertise, and they can be repeatable behaviors you activate on demand.

Anthropic says they can be used independently or together.

Wide reach, always on

  • Profile instructions: true everywhere, for you
  • Examples: your role, your units, the reading level you want, 'ask before assuming'
  • Changes how every chat begins

Narrow reach, on purpose

  • Project instructions: true for one body of work
  • Skills: a method you call on when a task needs it
  • Examples: one client's tone, a weekly report format, a review checklist

A sorting question for any rule

When you want to write a rule down, ask three questions in order.

  1. 1

    Is this true in every conversation I ever have?

    If yes, it is a profile instruction. "I work in finance and read metric figures." "When I ask for a summary, put the conclusion first."

  2. 2

    Is it true for one client, course or project?

    If yes, it goes in that project's instructions, and the supporting files go in its knowledge. "This client dislikes exclamation marks."

  3. 3

    Is it a method I use across several projects?

    If yes, it is a skill candidate: a checklist, a report layout, a way of reviewing. It travels with you instead of living in one project.

If a rule fits none of them, it is probably a one-off, and it belongs in the message you are writing.

The leak problem

The most common mistake is putting a project-sized rule in your profile. "Always write in a warm, informal tone" feels like a personal preference until the day you need a contract summary, and Claude's reply opens with a friendly greeting. Because profile instructions apply everywhere, they should be rules you would never want to turn off.

The test

For each profile instruction, ask: would this be wrong for any task I do in Claude? If you can think of one, move the rule down into a project or a skill.

The reverse mistake is also real. Retyping a genuinely universal rule in every project instruction block is how rules drift: you update one copy and forget the other four.

Memory is a fourth thing, and it is not written by you

One more mechanism carries context between chats, and it is different in kind. Memory is built by Claude from your conversations, and you edit it after the fact, where profile and project instructions are written by you in advance. Treat them as complements. Write the rules you are sure of, and let memory pick up the details you would never think to list, then review them from time to time.

One note of caution about Anthropic's help pages: the personalization article says project instructions are for paid plans in one place and says projects are available to everyone, including free accounts, in another. If you are on a free plan and a feature seems missing, check your own settings, not this article.

Writing instructions that Claude can use

Instructions fail in predictable ways. A few habits help.

  • State the rule and the reason. Anthropic's prompting guidance says that giving the motivation behind an instruction lets Claude generalize. "Keep paragraphs short because most readers see this on a phone" beats "keep paragraphs short."
  • Say what to do, not what to avoid. "Write in plain sentences of fewer than twenty words" is stronger than "do not write long sentences."
  • Write at least one 'if' rule. "If a number is not in the files I gave you, say it is missing." Rules for missing information prevent the most expensive mistakes.
  • Keep it short. Ten clear lines beat two pages. Long instruction blocks contradict themselves.
Prompt
“

Review my profile instructions below. For each line, tell me whether it is a universal rule or something that belongs in a project or skill, and give a one-sentence reason. Then propose a shorter version of the universal part.

”

Claude is a good editor of its own instructions. Ask it what contradicts, what is vague and what is missing, then decide which changes to accept.

A worked example

A freelance consultant has two clients and one house method for weekly reports.

RuleWhere it goes
"I prefer plain language and I read British spelling."Profile instructions
"Client A is a hospital group: cautious tone, no promises about outcomes."Client A project instructions
"Client B is a start-up: direct, informal, short."Client B project instructions
"Weekly report layout: summary, three findings, risks, next steps."A skill used in both

Nothing is repeated, nothing leaks, and changing the report layout is one edit.

When to revisit

Review your instructions when your role changes, when a client ends, and any time Claude keeps doing something you keep correcting. The repeated correction is the signal. If you correct the same thing three times, write it down in the right place.

Compare how other assistants handle the same split: Writing ChatGPT Custom Instructions That Actually Work and Custom GPTs vs. Projects.

Official sources

Checked on October 24, 2026. Features, plans and names change often, so the vendor's own pages are the final word.

Related Tutorials

  • Productivity
    5 min read

    Claude Connectors: What They Can Read, Change and Reach

    How Claude connectors work, the difference between reading and acting, the permission controls, and a short checklist before you connect anything.

    Read article
  • Productivity
    5 min read

    Claude Skills: When a Skill Beats a Saved Prompt

    What a Claude Skill is, the four kinds you will meet, how Claude decides to use one, and how to choose between a skill, a project and a pasted prompt.

    Read article
  • Operations
    5 min read

    Claude vs Gemini When Your Work Lives in Google

    How Claude reaches Drive, Gmail and Calendar through connectors, how its drafts get back into Google Docs, and where Gemini's built-in access is easier.

    Read article