All you need is claude

There's a cottage industry telling you to keep a stable of AI tools — this one for planning, that one for code, another for review — and to switch between them like a tradesman picking tools off a belt. I think that's mostly noise. You can run the entire software lifecycle with one capable agent, and you should.

Context is the bottleneck, not capability

The expensive part of working with an AI isn't the model's raw intelligence — it's the context you've loaded into the conversation. Every time you switch tools you throw that away and start re-explaining the problem from scratch. The cost of tool-hopping is the context you keep abandoning at the door.

One agent across the whole lifecycle

Here's a normal day, all in one place:

• Analyse: talk through the requirements, poke holes, surface edge cases before a line is written.

• Design: draft an approach, weigh trade-offs, write it down as a short plan.

• Build: implement against that plan, with the analysis still in context.

• Review: have it critique its own diff — and critique it yourself, because you're the one accountable.

• Test & QA: generate tests, run them, reproduce bugs, verify the fix.

Every step inherits the context of the last one. The agent already knows why the code looks the way it does, so its tests and its review are sharper than any fresh tool starting cold.

When to reach for something else

I'm not claiming there's never a niche where another tool wins. There are. But the default should be to stay put and keep your context intact, not to bounce between a dozen tabs because a benchmark said one model was two percent better at a thing you're not even doing.

Pick one good agent. Learn it deeply. Run the whole lifecycle through it. The compounding value of unbroken context beats a slightly better tool every single time.