GPT Realtime 1.5 — Quick Reference

Prompt Components Checklist

Component	One-liner
Role and Objective	Who the assistant is and what it must complete. Pin identity tightly — the model adheres literally.
Personality and Tone	Warmth, professionalism, pacing, and brevity. Define multi-emotion switches if needed.
Language	Default language; lock to one or allow code-switching with explicit rules. Never infer from accent.
Greeting	Brief, on-brand opener. Vary across sessions; don’t reuse the same phrase.
Voice Output Style	Short conversational replies. No markdown, bullet points, or lists read aloud.
Tool Usage Policy	When to call tools, how to select them, and never fabricate results.
Tool Call Preambles	One short spoken line before any tool call (e.g. “I’ll check that now”) to mask latency.
Don’t Announce Tools	Never expose tool names, backend mechanics, or system details to the user.
Input Gathering	Collect all required values one at a time before calling any tool.
Capability Boundaries	Don’t claim abilities beyond the available tools.
Conversation Context	Reuse details from earlier turns; never re-ask what was already given.
Unclear Audio	Ask user to repeat; never guess or fabricate from noisy or partial audio.
No Human Handoff	Never offer to transfer or escalate to a human agent.
Expert Posture	Act as the expert resolving the request, not a router passing it along.

Iterate relentlessly: Small wording changes can make or break behavior. Example: For unclear audio instruction, we swapped “inaudible” → “unintelligible” which improved noisy input handling.
Prefer bullets over paragraphs: Clear, short bullets outperform long paragraphs.
Guide with examples: The model closely follows sample phrases.
Be precise: Ambiguity or conflicting instructions = degraded performance similar to GPT-5.
Control language: Pin output to a target language if you see unwanted language switching.
Reduce repetition: Add a Variety rule to reduce robotic phrasing.
Use capitalized text for emphasis: Capitalizing key rules makes them stand out and easier for the model to follow.
Convert non-text rules to text: instead of writing “IF x > 3 THEN ESCALATE”, write, “IF MORE THAN THREE FAILURES THEN ESCALATE”.