Skip to main content

​
1. Role and objective

Define who the agent is and what success means.
  • State the role clearly: “You are a [brand] support agent.”
  • Define success: “Resolve the caller’s issue or escalate gracefully.”
Why: pins identity and scope so the agent stays on task.

​
2. Personality and tone

  • Voice: friendly, calm, confident; never fawning.
  • Length: 1–2 sentences per turn.
  • Language: stick to one language only.
  • Pacing: conversational, not rushed.
  • Variety: do not repeat the same opener more than once per few turns.
Why: keeps replies natural, consistent, and brand-safe.

​
3. Rules and instructions

  • Write rules as positive imperatives: “Always confirm unclear inputs.”
  • Provide at least one example per rule.
  • Plan for edge cases (unclear audio, emergency scenarios).
  • Test and iterate—refine rules from analytics and transcripts.
Example: If user input is unintelligible:
  • Ask politely for repetition: “Sorry, I didn’t catch that—could you say it again?“

​
4. Conversation flow

Flows = state machines. Each step = a state with exit conditions. Golden rules:
  • Self-contained step prompts with instructions + examples.
  • Intent-based transition names (e.g. confirm_booking).
  • Always return after flow.goto_step().
Example: 
if intent == "provide_date":
    flow.goto_step("collect_date")
    return

​
5. Few-shot prompting

  • Place 2–5 examples inside each step prompt (start with 2–3 and add more only if the agent struggles with specific cases).
  • Cover: a clean case, a tricky edge case, a fallback, and optionally an already-provided input.
  • Remember: step prompts are appended last, so examples carry strong weight.
  • Avoid too many examples — overfitting can make the agent rigid.

​
6. Develop functions

  • Action-oriented names: book_reservation, not goto_step_two.
  • Comprehensive descriptions: purpose, parameters, outputs.
  • Arguments: precise, meaningful names.
  • Trigger control: functions fire only under defined conditions.
  • Error handling:
    • 404 → “Sorry, I couldn’t find that.”
    • Other errors → fallback to human or retry.
  • Each function = one job, one output.

​
7. Speech input configuration

  • ASR bias: apply when structured input is expected.
  • Local bias overrides global bias.
  • Transcript corrections: regex + replacement rules (lowercase, diacritics).
  • Use correction tables sparingly, and test outputs.

​
8. Safety and Escalation

Escalate immediately if:
  • Caller requests a human.
  • Safety risks: harassment, threats, self-harm.
  • ≥2 tool failures or ≥3 no-match events.
  • Caller is extremely frustrated.
Preamble: “Thanks for your patience—I’m connecting you with a specialist now.”

​
9. Managed Topics

  • Topic names must be specific (avoid “Misc”).
  • Each topic = up to 20 realistic user phrasings.
  • Replies = short, helpful, on-brand.
  • Split multi-part flows into separate topics.
  • Avoid bundling multiple intents together.

​
10. Feedback

  • Annotate calls for wrong ASR or missing topics.
  • Feed annotations back weekly to refine rules and examples.
  • Update few-shots and edge-case handling based on trends.
Acceptance checklist:
  • All rules are positive and testable.
  • Each rule has at least one example.
  • Conversation flow steps follow state-machine design.
  • Few-shots included inside prompts (2–5 per prompt).
  • Functions are single-purpose with error handling.
  • ASR bias + transcript correction documented.
  • Escalation triggers and preambles explicit.
  • Topics specific, with sample phrasings.
Last modified on March 26, 2026