// CONTENTS
I want to say something nice about vibe coding first, because it deserves it.
Describing what you want to an AI agent and watching it appear is a real thrill. For a prototype, a script or a weekend project, it’s hard to beat. I’ve done plenty of it.
But I’ve also hit the wall that everyone eventually hits, and I hit it a long time ago.
The Wall
Vibe coding works beautifully right up until the project gets big enough to matter. Then the cracks show. The agent forgets a decision you made an hour ago. A fix in one place breaks something three files away. You ask for a small change and get a confident rewrite. And nobody, human or AI, can tell you what the software is actually supposed to do, because it was never written down. It lived in a chat scroll that’s long gone.
Scaling vibe coding to larger projects is impractical. Not impossible, just impractical: you spend more time steering, re-explaining and repairing than building.
None of this is new. It’s the same failure I watched over and over in QA at Intel, years before AI agents existed. As I wrote in I Am CoreyX, almost every bug is a disagreement between what someone meant and what someone made. AI agents don’t fix that problem. They just make it happen faster.
The Fix Was Never the Hard Part
I knew the fix, because it’s the same one that works for human teams: write down what you’re building before you build it, agree on it, and then build to the document instead of the vibe.
So that’s what I did with my agents. Requirements first. Then a technical plan. Then a design. Then a numbered checklist of tasks, each one tied back to the requirements it satisfies. Then implementation, one task at a time, with tests and a commit for each finished task. Then a walkthrough of what was built, a changelog entry and a version.
It worked, and it worked really well. The trouble was that it was repetitive. On every project I found myself typing the same instructions, the same document templates, the same rules about approval and traceability. I was prompting the process into existence, over and over.
When you notice you’re doing the same thing by hand every time, you know what to do. You automate it. So I turned it into skills.
What SpecDevShip Is
SpecDevShip is a pack of 19 agent skills that wraps an AI coding agent in a spec-driven pipeline. It runs in three steps, and then it repeats:
- Spec. Say what you want. The agent sorts it into a light tier, for everyday features and fixes, or a full tier, for new apps and real architecture, and writes the plan. You approve it.
- Dev. The agent builds it task by task, commit by commit.
- Ship. The walkthrough, the changelog and the version are done for you.
Around that core are the skills real projects need. Root-cause investigation for bugs, so the agent finds the cause before it writes a fix. Analysis of existing material, like an old app or a stack of mockups. Product, design and art-direction roles. And approval gates that track what’s a draft and what’s approved, so the agent can’t quietly change a plan you’ve signed off on.
The Claude Code Edition is out now, for $49. Future agent editions are included at no extra cost, and there’s a free field guide to the workflow if you just want the ideas.
Who It’s For
SpecDevShip is for anyone building apps or games with AI agents who has run into the wall. Especially if you’re building something you’ll still be maintaining a year from now, when “what were we thinking?” is a question someone actually has to answer.
The Proof Is the Site You’re Reading
This website was built with SpecDevShip.
Its spec had 28 requirements, broken down into more than 350 precise requirement statements and nearly 300 testable acceptance criteria. The plan was 45 tasks. Every task named the requirements it covered, every one was verified by automated tests before it was checked off, and when a requirement changed, the change went back through approval before any code moved. By launch, the site had more than a thousand automated tests behind it.
None of that was me typing process into a chat window. It was the skills doing what I used to do by hand.
Stop Vibe-Coding (Mostly)
I’m not telling you to give up the vibes. Prototype freely. Explore. Throw things away.
But when it’s time to build something real, write it down first. Your agent will be better for it, and so will you, six months from now, when you need to know why something works the way it does.
That’s why I built SpecDevShip. Get it at specdevship.com.