Skip to content

Instruction Node

NodeType = Instruction (NodeParams union discriminator 6)

Attaches a named instruction string to the current interaction as a slot. A later Generate or Generate and Speak node can reference it inside a Mustache template via the {{#instructions}} list or the {{slot.<name>}} direct lookup.

The node itself does not call a model and does not modify the conversation history that the model sees. It only adds a kind=Instruction SlotComponent to the current turn's interaction data, which the projection stage picks up when rendering the Generate node's template.

Parameters

Prompt

Param Type Default Range Structural Localizable Description
instruction str (multiline) empty — — ✓ Text later nodes can insert into their prompts, as the slot. tag or through the #instructions list.

Output

Param Type Default Range Structural Localizable Description
output_name str empty — ✓ — Name later nodes use to read the Instruction (in their Input, or as the slot. tag). Empty = this node's name.

Exits

Each exit is a structural string field on the node's params; its value names the target node (empty = END).

Exit Param field Description
default default_exit Node to run next. Empty = end the turn.

Exit routes

Exit Condition
default Always taken after the instruction is attached.

Mustache exposure

Given a node named greeter with instruction = "Welcome the user.":

Template expression Renders to
{{slot.greeter}} Welcome the user.
{{#instructions}}{{name}}: {{text}}{{/instructions}} greeter: Welcome the user.

Multiple Instruction nodes in one graph each contribute one entry to the {{#instructions}} list, in the order they executed.

Diagnostics

When enable_diagnostics = true, this node contributes to debug_info.nodes[].diagnostics. See Turn Diagnostics JSON for the envelope.

Instruction records a single key, and no input, output or engine section.

Key Meaning Absent when
parameters.instruction The static instruction text this node injected. —

Whether that instruction reached the model is a separate question, answered by the downstream Generate node's input.prompt[] — an Instruction node that ran is not proof its text was rendered, since that depends on the consuming node's template.

Example

from copy import deepcopy
from tryll_client.graph import GraphDescription, InstructionParams, GenerateParams, Placement

greeter_baseline = InstructionParams(
    instruction="You are a friendly greeter. Welcome the user warmly.",
    default_exit="generate",
)

graph = (
    GraphDescription()
    .add_node("greeter", greeter_baseline)
    .add_node("generate", GenerateParams(
        template="{{slot.greeter}}",
        placement=Placement.BeforeUserAsSystem,
        default_exit="",   # empty = END
    ))
    .set_start_node("greeter")
    .set_default_model_name("Llama 3.2 3B Instruct (Q4_K_M)")
)
agent = client.create_agent(graph)

# Change the instruction without recreating the agent:
p = deepcopy(greeter_baseline)
p.instruction = "You are a stern librarian."
agent.change_params("greeter", p)

using namespace Tryll::Client;
using namespace Tryll::NodeParams;

InstructionParamsT ip;
ip.instruction  = "You are a friendly greeter. Welcome the user warmly.";
ip.default_exit = "generate";
InstructionParamsT greeterBaseline = ip; // keep a copy for mutations

GenerateParamsT gp;
gp.template_  = "{{slot.greeter}}";
gp.placement  = ::Tryll::Placement_BeforeUserAsSystem;
// gp.default_exit = ""; // empty = END (the default)

GraphDescription graph;
graph.AddInstruction("greeter",   std::move(ip))
     .AddGenerate("generate",     std::move(gp))
     .SetStartNode("greeter")
     .SetDefaultModelName("Llama 3.2 3B Instruct (Q4_K_M)");
auto agent = client.CreateAgent(graph);

// Change the instruction without recreating the agent:
InstructionParamsT mutated = greeterBaseline;
mutated.instruction = "You are a stern librarian.";
agent.ChangeParams("greeter", std::move(mutated));