Rebounder Tech Blog

Written by the people who actually run these systems in production.

How I Settled on Splitting AI Coding Tools

公開 読了時間 約2分執筆: Rebounder 開発チーム(当該システムの運用当事者)

※本記事にはアフィリエイトリンクを含む場合があります。内容は広告の有無に影響されません。

結論

Rather than consolidating on one AI coding tool, splitting roles across complex development, quick prototyping and everyday lookups keeps both cost and accuracy steadier.

The short version

Up front.

Use What I use
Serious, complex development Claude Code (the main tool)
Getting a simple app into shape quickly Codex
Everyday conversation, lookups, image generation Gemini
Cursor Handed over its role; little used now

In one line: Claude Code as the main tool for development, Gemini as the everyday companion.

Claude Code — the main tool for complex development

An AI coding tool that runs in the terminal, and it will carry intricate implementation a long way when handed to it. This became the main tool for development.

The weakness, honestly: there is a sense of cost piling up the more you use it. The capability is not in doubt, but “use it freely without thinking” is not how it feels.

Cursor — the previous main tool

To be clear, Cursor did not get worse. At the time it was straightforwardly the main tool.

What came later was strong enough, so its turn simply stopped coming. That is all.

Codex — getting something into shape quickly

Heavy, complex work goes to the main tool; this is for getting something small into shape quickly. If recommending one to start with, I think it is this.

Gemini — the everyday companion

Less development, more everyday. I use it for lookups and image generation.

Why they stay split

I do not consolidate because the accuracy and cost each use needs are different.

Handing intricate implementation to a strong tool earns its keep, but running lookups and short prototypes through the same thing means the spend stops being proportional to the difficulty. Using an expensive tool on hard work is rational; using an expensive tool on easy work is just expensive.

On deciding to switch

Looking back, switching always happened because what came later was strong enough, never because the previous one got worse.

That ordering is worth being conscious of. Once you start explaining a switch as “the old tool is bad”, the judgement is usually clouded. Migration costs something, so a justifying story becomes necessary.

よくある質問

Q1If picking just one to start with, which?

Choosing for approachability, start with something that gets a simple app into shape quickly. Tools that take on complex implementation are powerful, but cost piles up the more you use them, so making one your main tool before you know what it is for leaves the spend unpredictable.

Q2Why not consolidate on the most powerful tool?

Because the accuracy and the cost each use needs are different. Handing intricate implementation to an expensive tool earns its keep, but running lookups and short prototypes through the same thing means the spend stops being proportional to the difficulty.

Q3How do you decide to switch?

It has happened in the order "what came later was strong enough", not "the previous one got worse". Once you start explaining a switch as "the old one is bad", the judgement is usually clouded.

この記事の根拠

  • ドキュメントファイル

本文の主張は、上の記録に書かれていることだけです。運用しているリポジトリは非公開のため リンクは張れませんが、どのファイルの何行目を、どのコミット時点で見て書いたかは 記事ごとに残しています。推測で書いた箇所はありません。