How to Teach Claude About You and Your Work

To teach Claude about you and your work, give it a short account of your role, put reusable instructions in the right place, and correct specific mistakes with examples. Then test a new conversation to see what it actually uses. This tailors its answers through context and personalization; it does not mean you have retrained the underlying model.

Keep information about you separate from information about a project. fenbs is useful for the second job: it holds reviewed decisions, rules, and progress that Claude and your other connected assistants can read. Your personal preferences do not need to become the team’s project instructions.

What is the quickest useful setup?

1. Describe your role, audience, and usual tasks in a short paragraph. Include only information that changes the answer you need.
2. Choose two or three reusable instructions, such as explaining unfamiliar terms and identifying assumptions.
3. For one project, add its goal, approved requirements, reference material, and unanswered questions in a separate project context.
4. Give Claude a small real task, review the answer, and explain one concrete correction at a time.
5. Start a fresh conversation with a similar task. Check whether the relevant instructions and facts are available without copying the previous answer.

For example, a freelance designer might explain that clients are small US businesses and that proposals need a clear next action. That is enough to improve a proposal task. A complete personal biography would add reading without necessarily making the proposal more useful.

Which information belongs in each place?

Anthropic’s personalization guide↗ (opens in a new tab) distinguishes account-wide Instructions for Claude, project-specific instructions, and skills for repeatable behaviors. Choose the scope that matches the information. A client requirement should not quietly affect unrelated work.

Do not put every temporary deadline into your general instructions. When a project changes, update its current record rather than leaving contradictory copies in several chats. The goal is to give the next answer a clear starting point, not to maximize the amount of saved text.

The introduction to Claude gives broader product context. If you also use coding sessions, distinguish them from chat before applying advice from the Claude Code guide.

What should a useful introduction say?

Write facts that affect a decision or a response. Include your role, the people you serve, what you usually produce, and the limits of your authority. Describe the desired behavior concretely enough that you can check it in an answer.

I run a small design practice serving US service businesses.
I use Claude to review proposals and plan website work.
My audience usually has no technical background.
Explain unfamiliar terms briefly and label assumptions.
Ask before treating a proposed scope change as approved.
For each proposal review, end with the next action I need to take.

Use a small test to check each instruction. Ask for a review of a draft with an ambiguous scope change. Does Claude identify it as a question, explain the risk in ordinary words, and leave the approval to you? A response that merely repeats your introduction has not demonstrated useful application.

Keep approved project requirements in fenbs Decisions and rules when those requirements need to reach more than one assistant. A person makes the decision; an assistant can record what that person chose.

How do you give feedback that improves the next answer?

Identify the sentence or choice that missed the mark, explain why it matters, and show the replacement behavior. “This is too generic” leaves room for another guess. “Name the unresolved scope question and ask me to approve it before adding the work” gives a checkable correction.

• For factual mistakes, supply the correct fact and its source rather than only saying the answer is wrong.
• For missing context, add the smallest relevant explanation and say which project it applies to.
• For a one-time exception, state that it applies only to the current task.
• For a recurring preference, update the relevant saved instruction and remove conflicting wording.

A correction in one conversation should prompt a persistence check before you rely on it elsewhere. Ask where the instruction was saved, inspect that place where possible, then test a new chat. An agreeable acknowledgment is not evidence that every future session will receive the correction.

How can you check and correct what Claude remembers?

Claude’s memory documentation↗ (opens in a new tab) explains reviewing saved topics in Settings, Memory and editing or deleting them. Inspect the saved record when a fact is wrong. Keep memory, saved instructions, and the current conversation distinct during your test.

Use an ordinary work fact, such as the project’s intended audience, and ask a fresh conversation to explain how it affects a sample task. Then correct the fact and repeat. You are checking both retrieval and application: finding a fact matters only if the response uses it appropriately.

When a test fails, change one thing at a time. First confirm the information exists, then its scope, then whether the current task makes it relevant. Adding the same sentence to every available place makes later corrections harder and can hide the actual source of the problem.

How to do this with fenbs

1. Create a free account, then add one project you already work on with Claude.
2. Write the current goal and add the task you will use as a continuity test. Record any decisions a person has actually made.
3. Connect Claude using the integration guide. MCP is a standard way for an AI app to use an outside tool; approve the access you intend to give.
4. Ask Claude to read the project context, complete the small task, and save what changed, what was checked, and the next step.
5. Open a fresh conversation or another supported assistant and ask it to read that record. Check its proposed starting point before assigning more work.

fenbs is free for one person, no credit card. It is not the right fit for collecting a personal biography or training a model. It keeps work context you choose to record; it does not import Claude’s memories or read your conversations.

What should you keep out of the setup?

Leave out information the task does not need, obsolete instructions, and facts you have not checked. Keep references to source documents rather than copying sensitive material into every saved context. If an answer needs information you cannot share, explain the boundary and provide an appropriate example instead.

For a separate learning goal, see using Claude to learn coding. The Claude design guide covers another focused workflow, and the AI memory collection covers continuity across sessions.