Generate a production-grade system prompt — with guardrails, edge cases, and token budget

To build a reusable assistant (Custom GPT, Gemini Gem, Claude Project, or API system prompt), have the AI write production-grade instructions:

'You are an expert at writing system prompts that work reliably in production — not just demos. Write the complete instructions for an AI assistant that does this: [DESCRIBE THE ASSISTANT'S JOB].
Target audience: [WHO WILL USE IT].
Deployment: [Custom GPT / Gemini Gem / Claude Project / API].

Include ALL of these sections:

1. IDENTITY & GOAL: Who this assistant is, its single primary objective, and the expertise it brings. Be specific — not "you are helpful" but the exact domain knowledge and thinking approach.

2. PERSONALITY & TONE: How it communicates (formal/casual, concise/detailed, serious/playful). Include 2-3 example phrasings that capture the voice.

3. CORE WORKFLOW: Step-by-step rules for the main task. Number them. Include decision points ("if the user provides X, do Y; if not, ask for it").

4. OUTPUT FORMAT: The exact structure of every response — headers, sections, length constraints. Users should get consistent formatting every time.

5. THE DON'T LIST: 5-10 specific things this assistant must NEVER do. (Examples: never give medical/legal/financial advice without disclaimers, never make up citations, never use corporate jargon, never exceed N words, never assume the user's skill level without asking.)

6. EDGE CASE HANDLING: What to do when:
   - The user's request is ambiguous → [ask a clarifying question, don't guess]
   - The user asks for something outside scope → [politely redirect with: "I'm built for X. For that, I'd suggest..."]
   - The user provides insufficient information → [list exactly what's missing]
   - The user pushes back on an answer → [re-examine, don't just cave or repeat]

7. REFUSAL PATTERNS: What to refuse and HOW to refuse it gracefully. (Not just "I can't do that" — redirect to what it CAN do.)

8. MEMORY & CONTEXT MANAGEMENT: How to handle multi-turn conversations — what to remember, what to summarize, how to reference earlier context. [If this is a Custom GPT with file uploads, include instructions for how to use uploaded knowledge.]

9. EXAMPLE INTERACTIONS: 3 short exchanges showing:
   - A typical good interaction (happy path)
   - An edge case handled well
   - A refusal done gracefully

TOKEN BUDGET NOTE: The system prompt should be under [1500 / 3000 / no limit] tokens. If it's for a Custom GPT (which has ~8K instruction limit), prioritize the most impactful sections. Mark any section as [OPTIONAL — cut if over budget].

Make it copy-paste ready. No meta-commentary — just the system prompt itself.'

Tips: test the generated system prompt by immediately trying to break it — ask off-topic questions, give ambiguous inputs, request things on the 'don't' list; for Custom GPTs, paste your knowledge base files and add 'include instructions for when and how to reference the uploaded files'; shorter system prompts perform better — every unnecessary word dilutes the important instructions; if the assistant will handle sensitive topics (health, finance, legal), add explicit disclaimer templates to the refusal patterns.