Claude CodeでRoo Codeのように対話ベースで進めたい Claude CodeでRoo Codeのように対話ベースで進めたい

As AI-assisted coding tools proliferate, many developers working in VS Code are finding themselves pushed to migrate from Roo Code to Claude Code — whether due to licensing changes, model quality preferences, or team policy. The transition, however, isn't seamless. One friction point that frequently comes up is the loss of Roo Code's hallmark interactive, step-by-step development style.

Roo Code, like Cursor, is built as a VS Code extension and is designed around a conversational loop — the AI checks in with the user at key decision points rather than barreling through tasks autonomously. Claude Code, by contrast, is Anthropic's CLI-based agentic tool that defaults to a more autonomous mode of operation. It's highly capable, but that same autonomy can feel like a loss of control for developers who prefer to review and approve each incremental step.

The solution proposed in this article is elegantly simple: prefix your prompts with a designated string — in this case, "対話:" (the Japanese word for "dialogue" followed by a colon) — and configure Claude Code via a system prompt or custom instructions to treat that prefix as a signal to operate in a step-by-step, confirmation-seeking mode. The idea is that by establishing this convention upfront, users can toggle between autonomous and interactive behaviors depending on the task at hand.

This kind of prompt-level behavioral control is becoming an increasingly important skill as LLM-powered tools embed themselves into professional workflows. Rather than waiting for tool vendors to expose explicit UI toggles, developers are learning to shape agent behavior through careful prompt engineering. It's a pragmatic workaround — and arguably a more flexible one, since the same technique can be adapted to other tools with minimal modification.

The broader context here is the industry-wide push toward autonomous AI agents in software development. OpenAI's Codex CLI, GitHub Copilot Workspace, and similar tools are all trending toward completing complex, multi-step tasks with minimal human intervention. Yet many practitioners remain skeptical of full autonomy, preferring a "human-in-the-loop" model where the developer retains clear review checkpoints. The tension between automation and control is one of the defining challenges of AI-augmented development right now.

Claude Code itself was formally released by Anthropic in 2025 and has attracted significant attention as a terminal-integrated agent leveraging the Claude 3 model family. It competes and overlaps with tools like Cline and Roo Code in the VS Code ecosystem, and the community has been actively exploring hybrid usage patterns — using different tools for different phases of a workflow, or porting best practices from one tool's paradigm into another.

Ultimately, the key insight from this article is that the "best" AI coding tool is less about raw capability and more about fit with a developer's preferred working rhythm. Prompt-driven customization offers a way to close that gap without waiting for vendors to ship new features. As the agentic tooling landscape continues to evolve rapidly, sharing these kinds of practical configuration patterns has real value for the community.