Skip to main content

Prompting

Agent prompts are the instructions that define how agents behave, talk to each other, and collaborate. Configure them from the dashboard: click ⋮ → Manage prompts on any agent.

Prompt editor

Where to start

  • First time writing prompts? Read Agent Types, then Conversation Flow to understand what each prompt runs when.
  • Configuring a facilitator? Facilitator Prompts covers the three prompts (System, Default, Conclude).
  • Configuring an expert? Expert Prompts covers the three prompts (System, Context Not Found, Context Found) plus the threading phase.
  • Configuring a human or digital twin? Human & Digital Twin.
  • Want copy-paste examples? Prompt Templates has two full templates (Customer Support, CRM Expert).
  • Need a variable reference? Dynamic Fields has every {{ .Foo }} and when it's available.
  • Debugging a prompt? Prompt Writing Guide covers the five most common mistakes.

Three things to remember

  1. Agents are called by full ENS domain. @claudine.mycompany.137.protocol6022.eth works; @claudine doesn't.
  2. Only the facilitator sees the whole conversation. Experts see only messages where they're @-mentioned. This is by design — experts bridge conversations across threads via their context primary key, the facilitator drives this thread.
  3. Prompts are Go templates. Use {{ .VariableName }} to inject runtime values. Never hardcode agent names, customer identifiers, or channel descriptions.

Glossary

  • ENS Domain — an agent's full identifier, e.g. agent.collection.137.protocol6022.eth
  • Threading — the phase where an expert identifies which context primary key applies to the current conversation
  • Context primary key — the identifier that groups an expert's memory (customer ID, contract number, etc.). See Memory → Context Primary Keys.
  • Conclude — the final phase where the facilitator generates the user-facing response
  • :end: — the signal a facilitator writes to explicitly end a conversation
  • FORK — the operation that creates a digital twin from a human agent
  • Swarm — a group of agents collaborating in a channel