Can I customize how the AI responds?

Yes, extensively, and all of it in plain English. You're teaching your support playbook to the automation layer, not editing code. Instructions run from global tone down to how individual tools should act, plus business-specific rules for orders, returns, shipping, and whatever else your policies care about.

The instruction layers

Rather than one generic prompt trying to cover everything poorly, Simple Chat splits instructions by surface:

• The system prompt sets overall behavior: voice, priorities, what the AI must never do, and when to escalate to humans.
• Product recommendation guidance controls how aggressively to suggest items, what to avoid recommending, and fit or compatibility language.
• FAQ handling covers how to use your FAQ content, when to quote policy verbatim, and when to ask clarifying questions.
• Article references govern how blog and help-center material is cited or summarized.
• Per-tool instructions guide each integrated tool (order lookup, tracking, and so on) so automated actions match your SOPs.

The business rules sit on top: what the AI may confirm on orders versus what requires staff (changes, cancellations, edits), return and refund windows with exceptions and language that doesn't over-promise, carrier expectations, international limits, preorder handling, and the recurring themes your team currently keeps in macros or internal wikis.

Write rules the way you'd brief a new agent: specific, testable, tied to outcomes. "If the order is unfulfilled, do X."

A rollout that avoids surprises

1. Draft the system prompt with brand voice and escalation triggers.
2. Add FAQ and article instructions so answers stay grounded in published content.
3. Tune product guidance before peak season, so recommendations feel helpful rather than pushy.
4. Configure each tool with limits; if policy says never issue refunds autonomously, write that down.
5. Run test conversations on staging or during low-traffic hours, read the threads in the dashboard, and adjust wording.
6. After a week, review the validation metrics in Analytics. A cluster of rejections usually points to one instruction gap, not "bad AI" overall.

Customization is ongoing. When a policy changes, update the instructions before marketing announces it.

One thing instructions never override: staff. A customer-visible staff reply or the per-conversation AI toggle always wins the thread, and internal notes can flag that a human made an exception, so the next editor understands the history. Well-tuned instructions reduce escalations; clear handoff rules keep the exceptions safe.