Claude Custom Instructions: Preferences, Projects, and Rules

Set Claude custom instructions at the smallest scope that fits: account instructions for your general preferences, project instructions for recurring project requirements, and the current message for a one-time task. Then test whether the output follows the instruction rather than assuming saving it guarantees the result.

This matters when the same person works on several clients or a team uses several assistants. Tools like fenbs keep human-approved project rules in a shared record, so a preference from one person’s Claude account does not become the only place a work requirement exists.

How do you set the right instruction first?

1. Write the behavior you want in one concrete sentence.
2. Ask whether it applies to all your conversations, one project, or only this request.
3. Save general guidance in account settings or project guidance inside the intended project.
4. Remove any older instruction that directly contradicts it.
5. Start a new test conversation in the relevant place and inspect the result.

Claude’s personalization documentation↗ (opens in a new tab) currently calls its account-wide setting “Instructions for Claude.” Open Settings from your account menu to find it. Use it for broadly useful preferences such as explaining unfamiliar terms, rather than the changing details of one assignment.

For example, “When explaining software, define a technical term the first time you use it” can sensibly apply across projects. “Always describe the client’s spring campaign as approved” belongs nowhere unless that approval is current, relevant, and documented for that project.

What belongs at account, project, or task scope?

A freelance designer might prefer short explanations personally, need detailed accessibility reasoning for one client, and request a long comparison for today’s decision. Those instructions are easier to understand when each states its purpose. Avoid an account-wide “always be brief” rule that makes necessary project explanations harder to produce.

Keep sensitive project information out of broad preferences. Even when an instruction is relevant, ask whether it needs a person’s or customer’s identifying details. A clear process rule can often be written without them.

When a requirement belongs to the whole team, record its approval in fenbs Decisions and rules. That gives connected assistants a common starting point beyond the settings of one account.

How do you add instructions to a Claude project?

Open the intended project and use “Set project instructions,” then save the guidance. Anthropic’s project setup guide↗ (opens in a new tab) says these instructions apply to chats within that project. Check that the conversation you are testing actually belongs there.

Keep instructions and reference material distinct. An instruction says what to do: “Use the approved terminology and flag unsupported claims.” A reference document supplies facts: a product description, research notes, or an accepted draft. Label those roles so a quoted example does not accidentally become a command.

Do not assume that an uploaded document’s mere presence settles every conflict. Point Claude to the relevant source for the task and specify what to do if sources disagree. A useful rule might be “If the current product sheet conflicts with an older draft, flag the difference before rewriting the claim.”

For broader workspace choices, see Claude Projects compared with custom GPTs. If you are working in the coding tool instead of a Claude conversation, use the Claude Code introduction to choose the appropriate instruction mechanism.

What does a clear instruction look like?

A clear instruction names a situation, an action, and a way to recognize success. “Be accurate” expresses an aspiration. “When a source does not support a claim, list the missing evidence instead of supplying a number” describes the next action.

Audience: small-business owners who are new to the topic.
Use plain US English and explain necessary specialist terms.
Use the supplied sources for factual claims.
Preserve qualifications such as may, estimated, and subject to approval.
If sources disagree, show the conflict before choosing wording.
Do not create quotes, customer examples, or statistics.
Separate the proposed draft from questions that need the editor’s answer.
For publication, a person must approve the final version.

Add a short before-and-after example when a rule is easy to misunderstand. If “estimated delivery” repeatedly becomes a promised delivery date, show both sentences and explain that the qualifier must remain. Examples should clarify the rule rather than introduce fresh facts to reuse.

Keep the instruction focused on behavior. A project biography, every previous correction, and a full archive of drafts will make the important requirement harder to find. For repeatable task procedures, Anthropic’s skills overview↗ (opens in a new tab) explains the separate role of task-specific guidance.

How do you test conflicts and missed instructions?

Choose a small test with a visible pass condition. Supply a paragraph with an estimate and ask for a rewrite. The draft should preserve the estimate and avoid turning it into a promise. Run a second test with a missing fact and check whether Claude flags the gap.

• Confirm that you saved the change in the account or project you intended.
• Check the latest request for wording that asks for a conflicting result.
• Look for vague requirements that need a specific action or example.
• Compare the actual output with the rule, including exceptions and qualifications.

Revise one problem at a time and keep the test input unchanged while comparing versions. Otherwise, a better answer may come from easier source material rather than clearer guidance. Record what failed and which revision helped so you do not rebuild the same test next month.

Instructions guide behavior; they do not provide permission enforcement or guarantee factual correctness. Keep human review for work that requires it. For another app’s equivalent settings, see ChatGPT custom instructions; the AI prompts collection covers the wider practice.

For guidance specific to other tools, see NotebookLM custom instructions, Perplexity custom instructions, and the custom GPT instructions template.

How to do this with fenbs

1. Create a free account and add the project whose instructions you want to share.
2. Record the person-approved work rules in Decisions and rules, keeping personal preferences out of the team requirements.
3. Save the reusable instruction prompt and a small pass-or-fail example in Skills and prompts.
4. Connect Claude or another supported assistant and ask it to read the project’s AI context before working.
5. Save test findings and the remaining next step on the task so the next session can continue from the accepted guidance.

fenbs does not edit Claude’s instruction settings or import your conversation history automatically. If you only need a personal answer preference, Claude’s settings may be sufficient. fenbs helps when project requirements and tested prompts need to remain available across people and supported assistants.