Custom instructions
| Field | custom_instructions |
| Type | Multi-line text (up to 4,000 characters) |
| Default | None |
| Required | No |
| Affects | Response generator |
Custom instructions are injected directly into the response generator's system prompt, in a clearly delimited section. This is the most powerful configuration option — and the most likely to cause problems if used carelessly.
Custom instructions are sandboxed with a disclaimer that platform safety rules take precedence. However, poorly written instructions can still degrade response quality. Read this entire page before using this feature.
What custom instructions are for
Use custom instructions for specific behavioral rules that cannot be expressed through the other configuration options. Examples:
Good uses:
Never recommend competitor products by name.
When a customer asks about our premium membership, mention the 30-day
free trial before discussing pricing.
If a customer mentions they are a returning customer, thank them for
their loyalty.
Always suggest the customer check our size guide before purchasing
clothing items.
Bad uses (avoid these):
❌ Be friendly and helpful.
→ Use the brand_voice field instead.
❌ Respond in Spanish.
→ Use the primary_language field instead.
❌ Keep responses short.
→ Use the response_length field instead.
❌ You are a helpful AI assistant for our store.
→ This is redundant — the system prompt already establishes this context.
How to write effective custom instructions
Be specific and actionable
The AI follows concrete rules better than abstract guidance.
| Vague (bad) | Specific (good) |
|---|---|
| "Be careful about returns" | "When a customer asks about returns, always confirm the return window (30 days) and mention that the item must be unworn with tags attached" |
| "Upsell when appropriate" | "When a customer is looking at a single product, suggest the matching accessory if one exists in our catalog" |
| "Handle complaints well" | "When a customer expresses frustration, acknowledge their feelings before providing a solution. Use phrases like 'I understand that's frustrating' or 'I'm sorry you're dealing with that'" |
Use positive language
Tell the AI what to do, not just what not to do.
| Negative only (less effective) | Positive + negative (more effective) |
|---|---|
| "Don't discuss pricing of competitors" | "Focus on our product's value when answering pricing questions. Do not mention competitor pricing." |
| "Don't make promises about delivery dates" | "Provide estimated delivery ranges based on the shipping information policy. Do not commit to specific dates." |
Keep it concise
Each instruction should be 1–2 sentences. The more text you add, the more the AI has to balance competing priorities. Fewer, clearer rules produce better results than a long list of detailed instructions.
Test after changing
After updating custom instructions, run 5–10 test conversations that exercise the new rules. Check:
- Does the AI follow the instruction?
- Did the instruction break anything else? (e.g., a rule about returns accidentally affecting exchange discussions)
- Is the AI interpreting the instruction literally in ways you did not intend?
What custom instructions cannot do
- Override safety rules. The Critic/Supervisor agent validates every response before delivery. Custom instructions cannot disable this.
- Access external systems. You cannot instruct the AI to call APIs, send emails, or perform actions outside the conversation.
- Change the AI's fundamental capabilities. You cannot make the AI perform tasks it is not designed for (image generation, code execution, etc.).
- Override the platform base prompt. The core identity, safety guardrails, and pipeline behavior are immutable.
Interaction with other settings
Custom instructions are applied after all other configuration (brand voice, formality, response length, policies). If a custom instruction contradicts another setting, the instruction wins for the specific behavior it describes, but the other settings continue to apply to everything else.
Example: If your response length is concise but your custom instruction says "When answering sizing questions, provide the full size chart," the AI will give detailed sizing answers but keep other responses concise.
© 2026 Remaker Digital, a DBA of VanDusen & Palmeter, LLC. All rights reserved.
Was this page helpful?