Claude custom instructions: how to make AI write in your voice
Custom instructions tell Claude who you are, how you write, and what to avoid. Step-by-step setup guide with real examples for business use.
Claude custom instructions are persistent settings you configure once and that activate automatically in every subsequent conversation. They tell Claude who you are, how you write, and what you never want to see in its output. The result: Claude consistently writes in your voice — without you having to re-explain it every time.
Without custom instructions, Claude has zero context about you. It writes for a hypothetical average user: formal, cautious, with lengthy introductions and closing summaries. The AI prose that every reader recognises immediately. Custom instructions solve that — once set up, permanently active.
What are custom instructions?
Custom instructions are a setting in Claude that lets you give the model persistent context about you and your work. In Claude.ai, find them under Settings → Profile → Custom instructions. In Claude Enterprise, you can set them at the personal level and at the project level — project-level instructions override personal ones where they conflict.
The difference from a regular prompt: a prompt is per-conversation and disappears when you start a new chat. Custom instructions are always-on background context — active in every conversation without you thinking about it. Claude reads them as part of its starting instructions for every session.
According to Anthropic, users who configure custom instructions are more than twice as likely to return to Claude regularly compared to users who rely on one-off prompts. The reason is straightforward: output that matches your own style and context is actually usable output.
Why it makes a difference
The impact of custom instructions is clearest in a direct comparison. Same question — "Write a recommendation email for our new service" — with and without instructions.
| Without custom instructions | With custom instructions |
|---|---|
| Dear [name], I hope this email finds you well. I am writing to bring to your attention an exciting opportunity that... | Hi [name], quick one: have you heard about our new approach to [problem]? We're helping companies like [company] to... |
| Furthermore, I would like to highlight that our service offers a broad range of capabilities that align with your needs. | Give me three minutes and I'll show you whether it's relevant. When works for you? |
Same Claude. Same question. Context makes the difference. With the right custom instructions, Claude writes consistently in the tone that fits your role, your organisation, and your audience — without you repeating yourself.
For organisations with multiple employees using Claude, there is an additional dimension: consistency. If ten colleagues each steer Claude in their own way, you produce ten different writing styles. With shared project instructions in Claude Enterprise, you create one organisational voice — across all conversations, all team members.
The four components
Effective custom instructions consist of four blocks. Each block serves a different purpose. Together they give Claude the context to write in your voice consistently.
- Who I am — name, role, organisation, and the audience you write for. Claude needs this context to hit the right level and tone. An IT manager at a construction company writes differently from a marketing director at a consultancy. Give Claude both.
- How I write — your language preference, tone (direct/formal/informal), typical sentence length, and ideally a reference document or example sentence: "Write like I do: short sentences, active verbs, no bullet soup, no introductions."
- What you never do — an explicit prohibition list of writing patterns. This component has the most impact. Typical items: "Certainly!", "As an AI...", bullet lists for single ideas, summarising what the user just said, overly formal sign-offs, vague closings like "I hope this helps!"
- Data rules — when Claude should pause and ask about anonymisation before continuing. Example: "If I mention a person's name, ID number, or unpublished financial figure, always ask whether I want it anonymised first." One sentence that structurally reduces unintentional data exposure.
A practical starter template:
I am [name], [role] at [organisation]. I write for [audience].
Tone: [direct / formal / informal]. Language: English.
Sentences: short. Active verbs. No introductions, no closing summaries.
Style reference: [link to existing document or example sentence].
Never use: "Certainly!", "As an AI...", summarising what I just said,
bullet lists for a single idea, "I hope this helps!"
Data rules: if I mention a name, ID number, or financial figure, always
ask whether I want it anonymised before continuing.Fill this in for yourself. It takes twenty minutes. The effect is permanent.
Examples by role
The same four blocks look different depending on your role and context. The table below shows concrete examples.
| Role | Key instructions | What Claude does differently |
|---|---|---|
| Operations manager | Short. Options in rows, not prose. Always include cost and timeline. | Gives three options with bullets instead of explanatory paragraphs. |
| Marketing manager | Write engagingly, not corporate. No passive sentences. CTA at the end. | Emails sound human and activating instead of generic. |
| Financial controller | Numbers always in tables. State assumptions explicitly. No rounding. | Analyses automatically include an assumptions list and exact figures. |
| IT manager | Technical terms are fine. Don't explain what an API is. Say what's broken. | Direct diagnosis instead of careful circumlocution. |
| Director / executive | Executive summary first. Max three points. No appendix mentality. | Every output starts with the conclusion, not the build-up. |
Note: instructions are cumulative. If you also add a style prohibition list (component 3), both apply. Claude weighs all active instructions simultaneously — there is no priority order unless you specify one explicitly.
Anti-AI writing style: the fifth layer
The best custom instructions include an extra layer: an explicit list of patterns that are immediately recognisable as AI output. The moment a reader spots that, your communication loses authority — even if the content is correct.
The most common AI writing patterns to explicitly prohibit:
- "In today's rapidly changing world..." — any introduction that opens with context instead of the message.
- The three-part structure as a crutch: "First... second... finally..."
- Sentences ending with "...and this is just the beginning."
- Bullet lists of five points where two would do. Padding for the sake of form.
- "It is important to note that..." — a sentence that never adds information.
- Vague sign-offs: "I hope this helps! Let me know if you have any questions."
- Alliterative section headers like "Results. Reflections. Recommendations."
Add this list as a fifth block in your custom instructions: "Always avoid the following patterns: [list]." Claude applies this consistently — the model recognises the patterns as explicitly prohibited and avoids them even when they might otherwise seem natural.
Common mistakes
Custom instructions only work when they are concrete and complete. The most common mistakes that undermine the effect:
- Too short. "Be direct and concise" is not an instruction. "Use a maximum of three sentences per paragraph, no introductory sentences" is an instruction.
- Too vague. "Write professionally" says nothing. Describe what professional means to you with a concrete example.
- Generic. Copy-pasting a template without your own context. Claude needs your specific role and audience, not an abstract description.
- Never updating them. Your writing style, role, or organisational context changes. Custom instructions need to keep pace. Schedule a brief review every quarter.
- Only negative. Prohibition lists work, but positive instructions are equally important. Tell Claude what you DO want to see, not just what you don't.
- Forgetting project-level instructions. Personal custom instructions do not apply in Claude Enterprise Projects. Set project instructions too when working in a shared workspace.
Custom instructions and Claude Enterprise Projects
In Claude Enterprise, custom instructions work at two levels: personal (for all your conversations) and per project (for a specific shared workspace). Project instructions take precedence over personal instructions within that project.
Practical use: set personal instructions for your own writing style and data rules. Set project instructions for the specific context of that project — the client, the product, the style guide for that engagement. This way Claude always writes in your voice and always in the right project context.
For teams using Claude Enterprise as their standard working environment, it makes sense to manage project instructions centrally. One person maintains the project context; all team members benefit automatically. This is the foundation of a consistent AI-driven communication style across an organisation — not per employee, but as a shared standard.
More on the full rollout strategy for Claude Enterprise — including the adoption plan per phase — in our article on the Claude Enterprise rollout guide.
Where to start
Setting up custom instructions takes twenty minutes. The effect is permanent. The best approach: start with one document you have already written that represents how you communicate well. Re-read it. What are the patterns? Short sentences or long? Active or passive? Formal or direct? Note that down and fill in the four blocks.
Then test with three concrete tasks you normally use Claude for. Compare the output with what you got before the instructions were in place. If the output better matches your style, the instructions are working. If it still sounds too generic, the fourth block — the prohibition list — is incomplete.
Working with Claude Enterprise and want to roll out custom instructions across your organisation? We build the full instruction set — personal level, project level, per department — as part of our Claude Enterprise adoption programme. Describe your situation via the contact form. We respond within one business day.