System prompt

• You need consistent voice, safety posture, and escalation rules across features.
• Swapping models or providers and want identical behavior without rework.
• Auditing responsibility for harmful output and need a single source of truth.

The system prompt is your policy surface: it encodes compliance stance, tone, and allowed tools. A tight system prompt reduces per-feature drift and shrinks QA scope when models change. PMs balance strictness (less risk, more refusals) versus flexibility (fewer user blocks, more variability).

Keep the system prompt short, structured, and versioned. Separate non-negotiables (safety, brand tone, escalation) from feature instructions that belong in user prompts. Add explicit refusal and redaction patterns. Run a daily smoke test suite that checks for regressions after model or content updates. In 2026, pair the system prompt with organization-wide style tokens (e.g., clarity level, empathy level) rather than prose paragraphs.

A budgeting copilot’s system prompt sets persona (“calm, pragmatic analyst”), redaction rules for bank data, and a strict refusal matrix. After upgrading to a faster model, refusal accuracy stays at 98% and CSAT holds at 4.6/5 with no new incidents flagged by trust & safety.

• Mixing user-level tasks into the system prompt, causing conflicts with downstream instructions.
• Letting multiple teams fork the system prompt without governance.
• Omitting explicit refusal language, leading to brittle safety behavior.