Spec-Driven YOLO Development
March 6, 2026
I’ve been having a ton of fun the last two weeks getting back into coding. I took a deliberate break the last 2 months to R&R after quitting Cocoon.
While I was gone, I was surprised to learn how many things changed in the AI coding tools world. When I left Cocoon, it was the dawn of Opus 4.5 and skills. After coming back, there are now new paradigms and tools like OpenClaw and Gas Town. Models are now bumped - Opus 4.6 and Codex 5.3 are the SOTA. I’m glad I’m not working right now so I can easily catch up with all of the wild changes going on!
One workflow I’ve been appreciating lately is using an AI agent + some kind of spec-driven development tool to help guide and flesh out the thing I actually want to build. After experimenting a little with a few tools, my favorite one right now happens to be Github’s spec kit. Combined with a container that can help me safely bypass permissions, I find I can “one shot” many prototypes with barely any guidance.
The workflow works something like this:
- Initialize spec kit, and chat with the workflow to define the kind of app or feature I want to build. Continuously refine the plan until it’s ready to implement. Use spec kit to break down the project into tasks.
- Tip: You can let CC come up with the technical stack but I think after letting CC choose it for me once, I realize it’s always going to be better dictating the tech you already know so you can fix things manually if need be. CC chose Rust + Tauri for a desktop application I’m trying to make, and I unfortunately don’t have any Rust experience right now.
- Spin up a container using
devc, which I’ve found the most braindead way to set up a containerized environment for Claude Code. - Set-up a multi-agent team that mirrors real-world development roles (back-end + front-end + QA + PM orchestrator + architect). Instruct the agent to spawn different agents with these roles and implement the feature.
- I use the multi-agent teams feature Anthropic recently shipped to help make this more natural (docs here).
- Let it rip! In ~30 minutes you get a prototype that’s maybe 70-80% there.
I’m sure it would be easier getting to 100% in fewer shots, due to possible skill issues on my part. But it’s been an absolute jam getting ideas for projects out of my head, and into a working E2E prototype somewhere.