I Tested Claude Code’s Ultracode. It Was Amazing, Until It Burned My Whole Week in Half a Day.

I tested Claude Code’s Ultracode mode because a couple of AI friends told me, “Alex, I built this thing in one night with a one-shot prompt and it worked.” Which, by the way, is a technique I never suggest. I wrote already about one-shot prompts, and every time someone says “just write the perfect prompt,” a small part of me wants to close the laptop.

But you know me. I’m always eager to test something when there is a chance that an AI tool can actually help me build faster. So I opened Claude Code, switched to Ultracode, gave it a very long epic with stories, and let it run. It was amazing for a few hours, then holy shit: in less than a day it consumed my whole weekly limit of the $200 subscription.

What Ultracode Actually Does

If you don’t know what Ultracode is, the simple version is this: it’s an agent orchestrator provided by Anthropic inside Claude Code, for users on at least the Max plan. So yes, starting from $200 per month. You give it what you need, it can create multiple agents, orchestrate them, run multiple cycles, maybe fix what they produced, and then come back to you.

The idea is genuinely good. I don’t want to make it sound like “terrible tool, don’t use it,” because that’s not the point. The orchestration idea is strong, but heavy, secret and unknown is expensive when you are not watching. You don’t really feel it while it is working, because it is doing a lot, and that is exactly the dangerous part.

The Expensive Lesson

I gave it a very long epic, with stories, something pretty big. So yes, I’m not innocent here. But burning the weekly limit in less than one day is not the kind of “wow” you want from a $200/month tool.

The first lesson is simple: the idea is good enough that I’m building my own version. Something lighter, like what I did with GSD, just translated into my k0 project. I don’t think all this heavy, secret, unknown stuff is always needed when the core idea can be translated into a skill.

An agent that can create other agents, but in a way that I can understand, control, and adapt. I prefer less magic if it means I don’t wake up with my subscription dead.

Iterate Fast, But Bring a Backup

The second lesson is something I already know, but I had another reminder: iterate fast and test fast. In this case I did it, so great, Alex. You need to learn anyway, and the fastest way to learn is to test it yourself.

The third lesson is maybe the most practical one: always have a backup plan. For me, that means another Anthropic subscription, because yes, I have one. Some friends keep saying, “Use ChatGPT as plan B.” I tried. But Opus works so well for me that I don’t want to move away just for a plan B.

A backup plan is not optional when your workflow depends on rate limits, token budgets, and models that can stop you in the middle of the good part. Maybe next time I will try Fable again with my k0 version of Ultracode, because I already had some thoughts on Fable vs Opus. Maybe Fable gets another chance.

So, Was It Worth It?

Yes and no. Ultracode showed me agent orchestration is going in the right direction. It also showed me that if the tool is too opaque, too heavy, and too hungry, I don’t want to depend on it.

The real takeaway is to build the smallest version of the idea that gives you control. Test the shiny thing, break something, pay the stupid tax, then build the version that fits my workflow.

If you are also playing with this stuff, drop a comment or get in touch. I’m always happy to talk with other AI geeks and understand what is working, what is not working, and what burned your subscription before lunch.