Define Prompt Skill

This skill is an action router — pick the step that matches the user's intent and execute only that step. Do not run other steps; do not parallelize.

Available actions:

• Step 1 — Single-string systemPrompt (default — when a one-paragraph instruction is enough)
• Step 2 — prompt { ... } builder with few-shot examples and structured turns
• Step 3 — Runtime context in a graph write session

Step 1 — Single-String systemPrompt

If the user just wants to set an instruction string, the factory's systemPrompt parameter is the right surface. No DSL needed:

import ai.koog.agents.core.agent.AIAgent
import ai.koog.prompt.executor.clients.openai.OpenAIModels
import ai.koog.prompt.executor.model.PromptExecutor

fun triageAgent(executor: PromptExecutor) = AIAgent(
    promptExecutor = executor,
    llmModel = OpenAIModels.Chat.GPT4o,
    systemPrompt = """
        You are a GitHub triage assistant. Classify issues, suggest labels,
        and link related issues by number when relevant.
    """.trimIndent(),
)

If the user is reaching for more structure than a single string supports, escalate to Step 2.

Finish here.

Step 2 — prompt { ... } Builder

Use the DSL when you need system + assistant + user turns interleaved (few-shot examples), or when the same prompt shape is reused across multiple agents.

import ai.koog.agents.core.agent.AIAgent
import ai.koog.agents.core.agent.config.AIAgentConfig
import ai.koog.prompt.dsl.prompt
import ai.koog.prompt.executor.clients.openai.OpenAIModels
import ai.koog.prompt.executor.model.PromptExecutor

val triagePrompt = prompt("issue-triage") {
    system("You are a GitHub triage assistant. Classify issues as bug, feature, or question.")

    // Few-shot example 1
    user("App crashes on Windows when I open the Settings dialog.")
    assistant("""{"classification": "bug", "confidence": 0.95}""")

    // Few-shot example 2
    user("Could we add dark mode to the export view?")
    assistant("""{"classification": "feature", "confidence": 0.9}""")

    // The actual turn is appended at runtime by the agent
}

fun fewShotAgent(executor: PromptExecutor) = AIAgent(
    promptExecutor = executor,
    agentConfig = AIAgentConfig(
        prompt = triagePrompt,
        model = OpenAIModels.Chat.GPT4o,
        maxAgentIterations = 50,
    ),
)

The Prompt class lives in ai.koog.prompt.Prompt as of 1.0 (moved from ai.koog.prompt.dsl.Prompt, #2022); the prompt { ... } builder DSL stays in ai.koog.prompt.dsl.

Pass a built Prompt through AIAgentConfig.prompt, as shown above. For per-node context, use Step 3.

Finish here.

Step 3 — Add Runtime Context in a Write Session

Add runtime data in the graph node where it becomes available. Pass the required data explicitly; append context before the next LLM request.

import ai.koog.agents.core.agent.context.AIAgentGraphContextBase

// Call addAccountContext(this, accountContext) inside the boundary node.
suspend fun addAccountContext(context: AIAgentGraphContextBase, accountContext: String) {
    context.llm.writeSession {
        appendPrompt {
            system("Account context: $accountContext")
        }
    }
}

Call the helper once at the intended boundary. Use user(...) for untrusted user-provided text instead of elevating it to a system instruction. For cross-session retrieval, invoke Skill(skill: "manage-state") Step 3. Its SystemPromptAugmenter / UserPromptAugmenter types belong to ai.koog.agents.longtermmemory.retrieval.augmentation and are configured inside LongTermMemory's retrieval { promptAugmenter = ... } block.

Write the selected prompt and agent configuration to the user's source file. Run ./gradlew build; verify that the model receives the intended instruction, few-shot turns or runtime context in the chosen agent/node.

Finish here.