Scaffold Agent Skill

Process steps in order. Do not skip ahead. This skill ends after Step 6 when the user has a runnable agent — do not chain into tool or MCP wiring; those are separate skills (add-tool, wire-mcp-server).

Step 1 — Clarify the Target

Ask the user, one question at a time:

• Target directory — must be empty or not exist yet
• LLM provider for the starter agent — OpenAI, Anthropic, Google, Ollama (local), or other (the user supplies the provider's Koog factory name)
• Build tool — Gradle Kotlin DSL (recommended) or Gradle Groovy DSL

If the user picks "other", confirm the factory exists in ai.koog:prompt-executor-llms-all (the umbrella). If not, plan to pull the specific prompt-executor-<provider>-client module — note this for Step 3.

Proceed immediately to Step 2 once all three are answered.

Step 2 — Verify the Target Directory

• If the path doesn't exist, create it
• If it exists, refuse to overwrite unless it is empty — silence is unsafe here; ask explicitly to confirm overwrite, and abort the skill on any answer that isn't "yes overwrite"
• Initialize git: git init in the target

Proceed immediately to Step 3.

Step 3 — Write build.gradle.kts

Use this template, substituting ${provider} placeholders:

plugins {
    kotlin("jvm") version "2.3.10"
    application
}

group = "com.example"
version = "0.1.0"

repositories { mavenCentral() }

dependencies {
    implementation("ai.koog:koog-agents:1.0.0")
    implementation("org.jetbrains.kotlinx:kotlinx-coroutines-core:1.10.2")
    testImplementation(kotlin("test"))
}

kotlin {
    jvmToolchain(17)
}

application {
    mainClass.set("com.example.MainKt")
}

If the user picked a provider that requires a non-umbrella client artifact, add that dependency explicitly (e.g., implementation("ai.koog:prompt-executor-litert-client:1.0.0")).

Write settings.gradle.kts with rootProject.name = "<directory-name>" and a single include(":") is not needed for a flat project — leave it as just rootProject.name = ....

Proceed immediately to Step 4.

Step 4 — Write Main.kt

Path: src/main/kotlin/com/example/Main.kt. Use the canonical 1.0 form — top-level AIAgent(...) factory, env-var API key, singleRunStrategy() is the default so no explicit strategy= parameter:

package com.example

import ai.koog.agents.core.agent.AIAgent
import ai.koog.prompt.executor.clients.openai.OpenAIModels
import ai.koog.prompt.executor.llms.all.simpleOpenAIExecutor
import kotlinx.coroutines.runBlocking

fun main() = runBlocking {
    val apiKey = System.getenv("OPENAI_API_KEY")
        ?: error("Set OPENAI_API_KEY in your environment")

    val agent = AIAgent(
        promptExecutor = simpleOpenAIExecutor(apiKey),
        llmModel = OpenAIModels.Chat.GPT4o,
        systemPrompt = "You are a helpful assistant.",
    )

    val result = agent.run("Hello! What can you help me with?")
    println(result)
}

For Anthropic, swap simpleOpenAIExecutor → simpleAnthropicExecutor, OpenAIModels.Chat.GPT4o → AnthropicModels.Opus_4_7, and the env var name. For Google: simpleGoogleAIExecutor + GoogleModels.Chat.Gemini_2_5_FlashLite. For Ollama: simpleOllamaAIExecutor(baseUrl = "http://localhost:11434") and an OllamaModels.* entry; no API key needed.

Do NOT add strategy = singleRunStrategy() to the constructor — it's the default and listing it muddies the example. Add an explicit strategy = ... only when the user is overriding the default.

Proceed immediately to Step 5.

Step 5 — Write Scaffold Static Files

.gitignore:

.gradle/
build/
.idea/
*.iml
.DS_Store

README.md (minimal — one paragraph, the run command, and the env-var requirement):

# <project name>

A Koog 1.0 agent. Set `<PROVIDER>_API_KEY` and run `./gradlew run`.

Proceed immediately to Step 6.

Step 6 — Hand Off

• Run ./gradlew --version to confirm the toolchain resolves. If Gradle isn't installed, point the user to https://gradle.org/install/ and finish here
• If Gradle is installed, run ./gradlew build once. If it fails, surface the exact error and the line in build.gradle.kts it refers to — do not auto-fix; coordinate dependency overrides come from the user
• If build passes, tell the user: "Set <PROVIDER>_API_KEY and run ./gradlew run to invoke the agent." Finish here.

Do not chain into add-tool or wire-mcp-server — those are separate user invocations. If the user immediately asks for a tool or MCP after this, those skills handle it.