Best Practices

This playbook highlights the tactics top SupportGPT teams use to maintain high-quality conversations at scale.

Refine the agent's instructions

• Set the tone: Clearly describe how the agent should greet, respond, and escalate.
• Define guardrails: Spell out topics the agent must avoid and when to hand off to humans.
• Add examples: Provide sample questions and answers to anchor the model’s behavior.

Revisit instructions whenever you launch a new campaign or product. Small tweaks can prevent confusing responses.

Improve readability of sources

• Convert complex PDFs to clean text or DOCX before uploading.
• Remove headers, footers, and watermarks that add noise.
• Keep content up to date. Replace outdated policies rather than uploading both versions.

Add suggestable links and media

• Maintain a list of approved resources (pricing, knowledge base, blog posts) and reference them in instructions.
• Use markdown links inside snippets so the agent can surface them naturally.
• Tag important URLs in your system prompt: Suggest https://example.com/faq for billing questions.
• Prepare short descriptions for key screenshots or diagrams so the agent can recommend them when relevant.

Choose the right model configuration

• GPT-4o Mini: Balanced accuracy and cost for most support teams.
• GPT-4o: High-stakes conversations requiring nuance and reasoning.
• GPT-4o Mini High: When latency matters more than perfect phrasing.

Adjust temperature thoughtfully:

Iterate with real conversations

1. Monitor escalations and unsatisfactory ratings in Usage.
2. Tag problematic transcripts and feed them back into your instruction updates.
3. Schedule weekly "agent tune-ups" with stakeholders to incorporate new learnings.

Utilize the Revise workflow

• Capture follow-up Q&A pairs when the agent misses the mark.
• Share revisions with teammates to align on messaging changes.