# Summary Most LLM failures come from lazy system prompts—teams write them once and forget them. The system prompt isn't a casual greeting; it's a job description that controls model behavior in production. Think of it like onboarding a brilliant but clueless new employee. You need to specify the role, behavioral constraints, exact output format, when to use tools, what the model should admit it doesn't know, and escalation rules. Vague directives like "be helpful and concise" with no format specification are basically useless. One ambiguous sentence—like telling a model to "include all relevant context" alongside JSON output—can waste a week of debugging because the model interprets "all relevant context" as stuffing extra keys into the response. The structure of your prompt matters measurably. Claude responds 15-20% better with XML tags (`<role>`, `<constraints>`, `<output_format>`), GPT models prefer markdown headers with JSON schemas, and Gemini handles markdown with numbered rules most consistently. Keep it atomic: one rule per line, not compound instructions buried in paragraphs. A 400-token system prompt outperformed a bloated 3,000-token version because the model got confused parsing everything at once. Split your content strategically—put role, hard constraints, and default output format in the system prompt (aim for 200-800 tokens), then inject task-specific instructions, documents, and user preferences per request. That's like environment variables versus function arguments. Five patterns actually work in production: make the role absurdly specific rather than generic ("billing support specialist for B2B SaaS with access to subscription tier and payment history" beats "helpful assistant"), use positive instructions over negative ones ("write in prose paragraphs" works better than "don't use markdown"), show exact output schemas instead of describing them in words (format compliance jumped from 70% to 95%+ with explicit JSON structure), add explicit uncertainty handling so the model admits when it doesn't know instead of hallucinating, and version your prompts like code with golden test sets to track compliance and success rates across changes.