AI Development is an "Improv Session"! Creating a New Groove for Three-Sword Engineers with Claude Code

AI development was supposed to be smooth. However, there was a "noise" that disrupted my development groove and almost made me lose sight of the product's core. But when I overcame that hurdle, development with AI dramatically transformed into an exhilarating experience, much like an intense impromptu session with a professional musician.

TL;DR

AI development with Claude Code can dramatically enhance consistency and efficiency not just through localized code modifications, but also through 'prompt engineering,' which tunes the overall product vision and architecture as an 'ensemble'.

Background/Environment: The Author's Triple Threat and Expectations for AI

Based in Osaka, I am a sole proprietor engaged in a "triple threat" career as a musician (head of Hormonize Records) × web engineer × business owner. In daily development, I use Next.js and TypeScript for the frontend, Python (FastAPI) for the backend, Supabase for the database, and GCP for infrastructure. Amidst the remarkable evolution of AI, I have constantly explored how to integrate AI into my development workflow to enhance overall productivity and creativity.

Among these, Anthropic'sClaude Codewas highly anticipated as my new development partner due to its excellent contextual understanding and fast responses. For me, who pursues a music style emphasizing tight rhythms and improvisation like BOOM BAP, interactive development with AI held the potential to create a true "new groove."

AI Development "Impromptu Session": Process and Early Groove

My AI development begins by presenting Claude Code with the product's business goals and a broad architectural overview. By providing detailed requirements, technical stacks, and desired functionalities, the AI takes charge of initial code generation and creating the skeleton of modules.

For example, when requesting the construction of a simple REST API with user management and authentication features, I start with a prompt like the following:

「あなたはPython (FastAPI) とSupabaseを用いたREST API開発のエキスパートです。ユーザー登録、ログイン、セッション管理、プロフィール更新機能を持つAPIを構築してください。セキュリティと拡張性を考慮し、モジュール分割も提案してください。」

Claude Code responds instantly, providing authentication logic, database schema suggestions, routing structures, and even test code templates. This initial interaction was an exhilarating experience, much like telling a session musician the theme of a song and having baseline and drum patterns instantly constructed.

I review the generated code and provide feedback on the architectural direction and specific business logic. Based on this feedback, the AI modifies and improves the code, and the development "groove" seemed to accelerate smoothly.

Stumbling Blocks and Solutions: Tuning a Disrupted Groove

The AI development, described as an 'improvised session,' wasn't smooth from the start. During a conversation with Claude Code, dependency conflicts and build errors frequently occurred in the local environment. Although the AI offered solutions each time, the overall vision I intended began to subtly deviate.

Specifically, when I requested a refactoring of a particular module, the AI focused solely on localized optimization, resulting in a loop where new bugs were created in conjunction with other modules. This 'technical noise' was like different parts playing out of sync during a recording session, creating dissonance. The 'groove' of development was disrupted, and I remember my focus drifting from the 'theme of the song' which was the essential value design of the product.

What puzzled me was that the solutions provided by the AI were too focused on immediate problem-solving. For example, there was an exchange like this:

筆者：「〇〇モジュールをリファクタリングして、テストコードを追加してほしい。」
AI：「承知いたしました。以下に修正案を提示します。（コードブロック）しかし、この変更は△△モジュールに影響を与える可能性があります。」
筆者：「では、△△モジュールも修正してほしい。」
AI：「承知いたしました。（コードブロック）この修正は〇〇モジュールと競合する可能性があります。」

In this way, solving one problem led to another, almost falling into an 'infinite loop' of debugging. This was precisely a situation where I was moving away from the 'musicality' of 'consistent, clean code' that I aimed for.

The solution was to stop coding for a moment.

Then, I re-presented the product's business goals and overall architectural design to Claude Code at a higher level of abstraction. Instead of localized correction instructions, I sought harmony as an 'ensemble' for the entire system. The specific prompt was as follows:

「一時停止しよう。現在の問題は局所的な修正の繰り返しから生じているようだ。この機能はプロダクト全体の『ユーザー体験のシームレスな流れ』という価値を提供する重要なセクションだ。現在のコードベースを、まるで一つの楽曲のように、各モジュールがそれぞれの役割を担い、全体として調和の取れた『アンサンブル』になるよう調律し直してほしい。特に、ユーザー体験の『リズム』を損なわないよう、パフォーマンスと保守性を重視してくれ。モジュールの依存関係を整理し、将来的な拡張性を考慮した設計を優先してほしい。」

Through this approach, Claude Code transformed its role from a mere code correction tool to a 'conductor' for the entire project. It was put back on track towards consistent, clean code that considered the independence and collaboration of each module. I felt this was exactly like an orchestra conductor pursuing not just the timbre of individual instruments, but the overall harmony.

Surprise and Learning: Establishing a New Groove for the Three-Sword Engineer

From this experience, the author was once again surprised by the depth of Claude Code's contextual understanding and its speed of response. There was an exhilarating feeling, as if improvising (playing improvisation) with a skilled musician. When the author presented abstract visions and concepts, the AI converted them into "notes" of concrete code and instantly began to "play." And if the author gave a slight correction, it immediately adjusted the "phrase." This high-speed feedback loop created a unique "groove" that is not obtainable in normal development.

Within this new workflow, the author's role also became clear. Routine tasks such as detailed implementation and debugging were left to the AI, allowing the author to focus on "conduction"—presenting the product's "vision" as a manager and conducting the entire "ensemble" of the codebase as a musician. This is precisely a new form of the three-sword style that the author pursues, and it is perceived as a significant step in workflow improvement in AI development. The author believes that prompt engineering is not a technique to manipulate AI, but rather "direction" and "communication" to maximize the potential of AI as an instrument.

Summary + Next Actions: Seeking a New Groove with AI

AI development is no longer a tool where humans unilaterally generate code. The author strongly believes it is an "improvised session" where humans and AI mutually stimulate each other and enhance creativity. Especially high-performance AIs like Claude Code can understand abstract concepts and make suggestions that consider overall harmony, allowing developers to focus on more fundamental value creation.

As the author experienced this time, even if "technical noise" occurs during development, by re-presenting the overall product vision and business goals to the AI and encouraging it to "tune" as an "ensemble", the "groove" of development will surely be regained.

This new style of AI development will significantly influence future development efficiency. The author will continue to pursue a "new groove" with Claude Code and advance more creative and fundamental product development. Readers, please experience "improvised sessions" with AI and find your own "new groove." The key lies not in letting the AI "conduct," but in becoming the "conductor" yourself and clearly presenting the vision through prompt engineering.