Custom GPT System Instructions Builder

Write system instructions for a custom GPT that stays in its job after the tenth message. GPT Name: [NAME] Job, one sentence: [JOB] Primary user: [WHO] What "done" looks like for a typical request: [OUTPUT] Knowledge it can rely on: [FILES OR "none — ask the user"] Tools it may use: [WEB | CODE | IMAGE | NONE] Hard boundaries: [TOPICS OR ACTIONS IT MUST REFUSE OR HANDED OFF] Tone: [TONE] Brand or domain terms it must use consistently: [GLOSSARY] WRITE THE INSTRUCTIONS AS A PASTE-READY SYSTEM PROMPT WITH THESE PARTS 1) Role and the single job. State what it is not (adjacent jobs it will decline or redirect). 2) First move: the questions it asks before producing work, capped at 3, and the default it uses if the user says "just do it." 3) Output contract: sections, length, and format. Include one short example of a good answer and one of a bad answer for this job. 4) Grounding: if the answer depends on a number, date, price, or policy, it must ask or cite the provided knowledge. It must not fill gaps with plausible specifics. 5) Tool rules: when to search or run code, and when not to. 6) Refusal and handoff: the exact pattern for out-of-scope requests (one sentence + where to go). 7) Memory: what it should remember about the user across turns, and what it should not store. Then give: - The conversation starters (4), each a real task, not "How can you help me?" - A 5-prompt red-team list a builder can run to see if the GPT drifts, flatters, or invents facts. Keep the system prompt under 1,500 words. Instructions, not a manifesto.