Spec vs Vibe Coding

Kiro offers two distinct modes for building applications. Understanding when to use each is key to getting the most out of the IDE.

Spec (Recommended)

Spec-driven development is Kiro’s recommended approach. When you provide a prompt in Spec mode, Kiro creates three structured documents:

1. Requirements — User stories and acceptance criteria in EARS notation
2. Design — Technical architecture, sequence diagrams, and implementation considerations
3. Implementation Plan — A detailed task list with discrete, trackable tasks

Use Spec mode when you want a structured, reviewable development process.

Vibe

Vibe coding is a more freeform mode. It’s great for:

• Rapid exploration and testing
• Refining prompts before using Spec mode
• Quick prototyping
• Building when requirements are unclear

As we’ll see later, you can combine both — use Vibe to refine your prompt, then pass the result to Spec mode.

Model selection and Autopilot

• Model selection — You can choose from available models, though Auto is recommended. Kiro will optimise which model to use. Model selection influences both results and credit usage.
• Autopilot mode — An autonomous execution feature where Kiro can make multiple code changes, run commands, and make architectural decisions across a codebase to complete complex tasks with minimal human intervention.

Tip

Using the Auto model setting lets Kiro pick the best model for each task, balancing quality and credit usage.

Open Weight Models

Kiro now supports open weight models in addition to Claude. These models offer different strengths and credit costs:

Model	Credit Multiplier	Best For
DeepSeek v3.2	0.25x	Agentic workflows, multi-step tool calling, maintaining state across long sessions
MiniMax 2.1	0.15x	Multilingual programming (Rust, Java, Go, C++, Kotlin, TypeScript), UI generation
Qwen3 Coder Next	0.05x	Long agentic CLI sessions, 256K context support, budget-friendly coding

Note

Open weight models are available to both free and paid users. Select them from the model dropdown in the IDE or CLI. Inference runs in AWS US East (N. Virginia) region.

When to use open weight models:

• Quick iterations and boilerplate — Use Qwen3 Coder Next (0.05x) for routine tasks
• Complex debugging — DeepSeek v3.2 excels at maintaining context across debugging sessions
• Frontend development — MiniMax 2.1 is strong at generating web, Android, and iOS UI code