Design the conversation

Design stages as business decision checkpoints, not as individual spoken turns.

The flow boundary

The model listens and speaks. Dino owns the graph and applies the model's reported outcome.

Caller ──▶ Model conversation ──▶ next(outcome, evidence)
                                      │
                                      ▼
                              Dino routes or ends

Create a stage only for one of these reasons:

1. A branch sends the call to different targets.
2. A gate must succeed before the call can continue.
3. A tool result changes the next action.
4. A dispatch selects one sub-flow.

Put ordinary questions, re-asks, and explanations inside the stage instruction. A focused objective usually needs two through five stages.

Choose the right data tool

Need	Use
Recall facts later in the call or post	memory
Write a structured result that the model does not read	capture
Call an external system	Agent-owned integration
Move, dispatch, or end the flow	System next tool

Prompt your AI agent

Recommended

Use the `dino-ping` flow-authoring guidance to design this conversation: <brief>.
Inspect F02 in the installed skill for stage density.
Inspect the nearest capability example for tools.
Treat stages as decision checkpoints, not turns.
Define explicit success and failure conditions for every stage.
Use memory only for facts needed later.
Use capture only for structured reporting.
Validate the full agent after each flow edit and render the final graph.
Do not publish or place a call.

Use the Dino CLI directly

dino agent validate --agent-dir <absolute-agent-dir>

dino agent render --agent-dir <absolute-agent-dir> \
  --output <absolute-agent-dir>/preview.png --overwrite \
  --confirm-mutation agent.render

Gotchas

• More than eight stages for one objective usually indicates over-decomposition.
• A route target must exist and remain reachable from the start stage.
• Every path must reach a terminal outcome.
• A rejected tool call stays in the current stage. It is not a failure outcome.
• Keep corrective guidance short and imperative.
• Allow each tool only in the stages or posts that need it.

Next

Use In-call tools for integrations, or Give it a voice for lifecycle audio.