System Prompts and Instructions

Audience: professionals who use Claude at work, no coding or API knowledge needed. Verified = checked against Anthropic's documentation or help center on 2026-10-16. Design guidance = widely taught practice, not a documented rule.

"System prompt" here means the standing instructions that apply to every chat, such as a Project's instructions (Anthropic's help center: project instructions tailor Claude's responses, for example "a more formal tone" or answering "from the perspective of a specific role or industry", verified).

What good instructions contain (design guidance, built on verified prompting principles)

1. Role and audience: who Claude is acting as, who it is writing for.
2. Purpose: what this Project or setup is for.
3. Rules: must do, must not do, in plain language, with the reason for important rules.
4. Format and tone: length, structure, style, terminology (house style).
5. How to handle uncertainty: say "I don't know", ask a question, or flag assumptions; never invent facts.
6. Sources: which uploaded documents are authoritative, and what to do when they conflict or do not answer.
7. Examples: one or two short examples of an ideal answer.
8. Boundaries: what to refuse or escalate (for example, anything involving personal data or legal advice).

Writing tips (verified from Anthropic's prompting guide)

• Be clear and direct; explicit beats implied.
• Use numbered lists when order matters.
• Give reasons so Claude generalises.
• Show examples (3 to 5 for best results where you can).
• Keep each instruction specific; contradictory instructions can make Claude choose arbitrarily.
• Organise sections with clear labels.

Keep them short and testable

Long, rambling instructions are followed less reliably. Aim for the fewest instructions that produce the behaviour you need, then test with realistic questions (D5.2).

Maintain them

Review when your policy, product or audience changes; remove outdated rules; keep a dated version and a note of why you changed it.

Do not rely on instructions for hard limits

Instructions guide Claude; they are not a security control. Do not put secrets in them, and do not assume they stop a determined misuse. Use organisation controls and human review for hard rules (D6.4).

Exam traps

• Vague instructions ("be professional").
• A long list of overlapping or conflicting rules.
• No reason or example for important rules.
• Never updating after policies change.
• Secrets or confidential data inside instructions.