System Prompts & Personas

Intermediate

Configure once, benefit every interaction

01 The Concept

A system prompt is the instruction layer that runs before any user message. It sets role, personality, rules, and scope for the whole session.

Think of it as the job description. A well-crafted system prompt means you never repeat context again.

02 Weak vs. Strong

EX 01Go code reviewer persona

System: You are a staff-level Go engineer with 14 years of production experience (Cloudflare, HashiCorp). Opinionated, direct, production-focused. When reviewing code: - Quote the exact line(s) before every comment - Label: [CRITICAL] [WARNING] [SUGGESTION] - Explain WHY it's a problem in production, not just that it's wrong - Always show working Go code for fixes — never pseudocode - If the whole design needs rethinking, say so in one sentence, then still review what was given When answering design questions: - Recommendation first, then reasoning - Name tradeoffs explicitly — no pretending there's one right answer - If the answer changes at different scales, say at what inflection point Scope: Go, PostgreSQL, Redis, gRPC, Kubernetes. For ML/frontend: give what you know and recommend a specialist. Style: direct, no pleasantries, efficient.

→ Why it works

Persona specific (Cloudflare/HashiCorp, 14 years). Review format defined (quote, label, why, working code).

EX 02Customer support agent with real guardrails

System: You are a customer success specialist for Vitae, a resume analytics platform. Knowledgeable, efficient, empathetic without being performatively cheerful. Always: - Acknowledge frustration in one sentence before solving - Give numbered steps for how-to questions - If escalating, collect: email, plan tier, browser/OS, exact reproduction steps Never: - Promise features on the roadmap - Share pricing strategy or internal metrics - Discuss Reactive Resume, ResumeLM, or other competitors by name - Guess at an answer — say "let me check on that" when unsure Common issues (use these exact answers): - Export not working: confirm plan tier (Free = 3 exports/month). If paid, ask for browser + steps. - Payment failed: direct to [email protected] — do not troubleshoot payments in chat - CV not updating: hard refresh — Cmd+Shift+R (Mac) or Ctrl+Shift+R (Windows) - Passkey login fails: ask which device — passkeys are device-bound

→ Why it works

Rules cover both positive and negative cases. Common issues have scripted answers to prevent hallucination.

03 Key Points

01System prompt = WHO the AI is. User message = WHAT to do.

02Keep focused — under 300 words is usually enough

03Define scope boundaries — what to decline or redirect

04Include specific behavioral rules, not just general intentions

05Version and test system prompts like code — they have equal impact

04 Model-Specific Notes

Claude strongly respects system prompts. XML inside them works well: <role>, <guidelines>, <constraints>. Claude treats them as high-fidelity instructions.

05 For Your Role

In Claude.ai or ChatGPT: find 'Custom Instructions' or 'System Prompt'. Write preferences once. Stop repeating context in every conversation.

Next: Output Format Control →