Skip to main content

​
Prompting & Agent Design Basics

Your agent configuration is the single biggest factor in how your calls feel. This guide covers the fundamentals of designing prompts, wiring tools and knowledge, and using context so your agents behave predictably in production.

​
Who this is for

  • Anyone configuring agents in the Truedy dashboard.
  • Developers writing prompts and integrating external tools/knowledge.

​
The four levers of agent behavior

Every Truedy agent is controlled by four main levers:
  • System prompt – What the agent should do and how it should talk.
  • Voice – How the agent sounds.
  • Knowledge bases – What static information it can reference (docs, FAQs, policies).
  • Tools – What actions it can take (look up a customer, book a meeting, write to your system).
When you tune an agent, you’re usually adjusting one or more of these.

​
Writing strong system prompts

Good prompts are:
  • Specific – Clear role and responsibilities.
  • Constrained – Clear boundaries of what the agent should not do.
  • Example-driven – A few sample interactions or bullet-point behaviours.
  • Tone-aware – Simple instructions about style: friendly, concise, formal, etc.

​
Template for a robust prompt

Use this structure when you start a new agent: 
You are a [ROLE] for [COMPANY]. Your goal is to [PRIMARY GOAL].

Tone:
- [e.g. Friendly and professional]
- [e.g. Use simple language. Avoid jargon.]

You MUST:
- [Rule 1]
- [Rule 2]
- [Rule 3]

You MUST NOT:
- [Don't make up policies or prices.]
- [Don't promise discounts or features that are not documented.]

When you don't know:
- [Explain what you can do instead]
- [Offer to escalate / collect details]
You can then layer in task-specific rules (e.g. for sales, support, collections, reminders).

​
Using context variables

Truedy calls accept a context object:
  • Keys and values are strings.
  • Your agent sees them as template variables (e.g. {{first_name}}, {{company}}).
Use them to:
  • Personalize greetings (Hi {{first_name}} from {{company}}).
  • Insert dynamic information (plan name, debt amount, renewal date).
  • Drive branching logic (different scripts per segment).
Combine this with a prompt section: 
You receive a JSON context object for each call.
Use these keys when present:
- first_name
- company
- segment
- last_order_amount

​
Designing agents for specific use-cases

​
Support agent

  • Goal: Resolve issues or route to the right place.
  • Key ingredients:
    • Rich knowledge bases (docs, FAQs, policies).
    • Clear escalation rules in the prompt.
    • Tools for ticket creation or CRM updates.

​
Sales / outbound agent

  • Goal: Qualify leads and book meetings.
  • Key ingredients:
    • Prompt that focuses on asking questions first, then recommending.
    • Calendar or CRM tools to actually book a meeting.
    • Batch calls / campaigns configured with good contact data.

​
Reminders / notifications

  • Goal: Deliver a clear message and possibly collect a simple confirmation.
  • Key ingredients:
    • Highly constrained prompt (few allowed branches).
    • Clear instructions for DTMF or verbal confirmations.

​
Tools and knowledge: when to use which

  • Use knowledge bases when:
    • The information is mostly static (docs, policies, product catalog).
    • You want better semantic recall in long-form answers.
  • Use tools when:
    • You need fresh, dynamic data (current balance, latest order, open tickets).
    • You want the agent to perform actions (book meetings, update records).
Most production agents use both:
  • Knowledge bases for background understanding.
  • Tools for real-time actions and up-to-date information.

​
Testing and iteration loop

  1. Start with a simple prompt and one main goal.
  2. Run several calls (or simulations) with real-ish data.
  3. Collect:
    • Where the agent hesitated or hallucinated.
    • Where it failed to follow rules.
  4. Tighten the prompt:
    • Turn implicit expectations into explicit MUST/MUST NOT bullets.
    • Add one or two example flows.
  5. Repeat until behaviour is stable.
Use the call details view and transcripts to see exactly what the agent did and adjust accordingly.

​
Next steps