Product GPT Prompts

InSkill’s AI copilot uses a layered prompting architecture. At the foundation, the platform includes a protected system prompt that governs model behavior and overall system functionality.

On top of that baseline, each product can define its own Product GPT Prompt — instructions written by product experts in plain language that allow product experts to shape response style, terminology, and safety emphasis without touching the underlying system prompt.

What Product Prompts Can Do
✅ Add product terminology / naming conventions
✅ Provide product-specific guidance rules (e.g., “never recommend adjusting X setting without Y pre-check”)

Unlike traditional LLM prompting — which can alter model behavior and introduce risk — Product prompts operate within InSkill’s guardrails:
🔒 The system prompt and retrieval logic remain protected
❌ Do not require AI or prompt engineering expertise.

How To Use

Best Practices
✅ Clear rules, simple language...shorter is better

⛔ Avoid
❌ No prompt syntax: “You are a…”
❌ "Format output as HTML/JSON/etc."

Formatting
Product GPT Prompts should be written in natural language...markdown can be used when structure improves clarity — such as lists, examples, or grouped guidance.

Terminology
- Force Sensor (not Load Cell)
- Drive Enable (not Motor Power)
- Heater Bank A/B (not Bank 1/2)

Safety
- Include: “Lockout/Tagout required”

🔄 Prompt Lifecycle Behavior

• Prompt updates apply when a new conversation is started
• Prompts are inherited by any copilot that uses that product as a component

Examples

📌 Terminology

• “Refer to the ‘Load Cell’ as Force Sensor in all operator-facing responses.”
• “Use the abbreviation MFI instead of Melt Flow Index.”
• "Oil = coolant"

⚠ Safety & Compliance

• “If the user must access internal panels, include: ‘Lockout/Tagout required.’”
• “If a step involves lifting heavy parts, remind them to use two-person lift.”

🧭 General Guidance

• “Use metric units unless the question uses imperial.”
• "Assume 5-6 digit numbers without additional context are alarm codes"
• “When recommending calibration steps, include the expected duration for planning.”