Most response quality problems trace back to one of three things: a weak prompt, missing knowledge, or the wrong model. This guide gives you a practical playbook for fixing each.
1. Strengthen the system prompt
The system prompt is the highest-leverage lever. A great prompt does four things:
- Names who the agent is and what it does
- Sets the tone and style
- Lists explicit "don'ts"
- Tells the agent how to handle "I don't know"
Add to your prompt when you see specific failures:
- Hallucination ("invents facts") → add: "Never make up information. If you don't know, say so honestly."
- Verbose ("rambles") → add: "Keep responses to 2–3 sentences unless asked for detail."
- Off-brand tone → add 2–3 example responses in the voice you want.
- Off-topic ("answers anything") → add: "If the question isn't about [scope], politely redirect."
2. Add or fix sources
If the agent answers the right kind of question wrongly, the issue is usually the source material — not the prompt.
- Out-of-date pricing → re-crawl the URL or upload the latest PDF
- Conflicting answers → two sources disagree. Delete the older one.
- Missing detail → add a Q&A pair for the exact question
- Returns the wrong policy → add a single, canonical policy as a text source so it dominates retrieval
3. Pick a stronger model
If you've hardened the prompt and the sources are clean but answers still feel weak, the model itself may be the bottleneck. Upgrade to a frontier-tier model for:
- Multi-turn reasoning ("if X then Y, but only when Z…")
- Sales conversations where nuance matters
- Complex troubleshooting
- Multilingual customer bases
4. Use feedback signals
Turn on Collect feedback in Settings → Interface. Customers click thumbs up/down on every agent reply. Review the thumbs-down responses weekly — they're a goldmine of prompt and source improvements.
5. The Playground iteration loop
See the Playground guide. The fastest path to a great agent is to spend an hour in the Playground every week, asking the questions your customers ask, and updating the prompt or sources whenever an answer falls short.