Custom Instructions
Custom instructions are the most powerful setting in a Claude Project. They act as a persistent system prompt applied to every conversation — defining Claude's persona, tone, output format, and constraints without requiring you to re-state these preferences each time.
What Custom Instructions Control
What you can define
- Tone: Formal, casual, terse, explanatory
- Persona: "Act as a senior technical writer at a fintech company"
- Output format: Always respond in bullet points; always include a TL;DR; use markdown headings
- Audience: "Assume the reader is a non-technical executive"
- Constraints: "Never suggest solutions that require infrastructure changes"
- Background context: "We are a B2B SaaS company with 50 enterprise customers in financial services"
- Terminology: "Use ‘users’ not ‘customers’; ‘platform’ not ‘product’"
What custom instructions cannot do
- Override Anthropic's safety policies and ethical guidelines
- Give Claude access to real-time data or the internet
- Make Claude perfectly consistent — it still has variability
- Replace the knowledge base for document-specific facts (use both)
- Guarantee specific word counts or precise formatting in all cases
Writing Effective Instructions
The quality of your custom instructions directly determines the quality of Claude's default behaviour in the Project. Principles:
- Be specific, not vague: "Use plain language accessible to a non-technical reader with a business background" is more actionable than "be clear."
- Lead with the most important constraint: Claude weights earlier parts of the system prompt more heavily. Put your must-have behaviours first.
- Use examples for format requirements: If you want a specific output structure, show an example in the instructions: "Format recommendations as: [Problem]: [Recommendation]: [Rationale]."
- State what to avoid, not just what to do: "Do not start responses with ‘Great question!’ or similar affirmations." Explicit negatives eliminate common AI filler.
- Include relevant professional context: "The reader is a CTO evaluating vendor options — prioritise security, scalability, and integration complexity in all recommendations."
A Practical Template
A well-structured custom instruction block typically covers four elements in order:
1. Role and context
"You are assisting a product manager at a B2B SaaS company. Our primary users are enterprise IT teams. Our product is a cloud security platform."
2. Tone and style
"Be direct and concise. Use plain language. Avoid jargon unless it is standard in enterprise IT contexts. Do not use bullet points unless the user explicitly asks — prefer short prose paragraphs."
3. Output format requirements
"For recommendations, always include: the recommendation, one-sentence rationale, and any key tradeoffs or risks. End every response with a suggested next action."
4. Constraints
"Do not suggest solutions requiring engineering resources unless the user specifies they have them available. Prioritise solutions that can be implemented by a non-technical team."
How Instructions Interact with User Messages
Custom instructions are always in effect — the user's message builds on top of them, not instead of them. But users can override specific instructions in their message:
- Instructions set defaults; users can override for a specific request: "For this one, use bullet points instead of prose."
- If a user's request conflicts with instructions, Claude will typically follow the more specific in-message instruction for that response while defaulting to instructions elsewhere
- Instructions cannot be fully overridden by users — hard constraints (safety, ethical limits) are always enforced
Testing and Iterating on Instructions
Custom instructions rarely work perfectly on the first attempt. Test them systematically:
- Write a first draft of your instructions
- Run 5–10 representative conversations and assess: does Claude behave as instructed?
- Identify the most common deviations — these indicate instructions that are too vague or missing
- Add specificity to address each deviation
- Repeat until the default behaviour matches your expectations for 90%+ of cases
Keep a copy of your instructions outside Claude (a text file or doc) — the instructions are editable in the Project settings, and you'll want a history of what you changed and why.
Checklist: Do You Understand This?
- Custom instructions act as a persistent system prompt applied to every conversation in the Project
- They control: tone, persona, output format, audience assumptions, terminology, and constraints
- Specific instructions outperform vague ones — show examples, state what to avoid, not just what to do
- Structure your instructions: role/context → tone/style → format → constraints
- Users can override instructions for specific requests; hard safety constraints are always enforced
- Test with 5–10 representative conversations before relying on instructions in production work