Forty-five minutes on a spec. Thirty minutes to build.
Visionaire Specication System - Claude Code Plugin
Thirty-eight AI personas talking to each other in a desktop app I had not written a single line of code for.
They had opinions. They interrupted. You could configure how much of each trait any one of them brought to a conversation: creativity, directness, analytical depth. You could add more of them to the room mid-session. They’d all start talking.
I had built this application before lunch. Not by coding it. By designing it. Six phases of structured interviews, decisions, and reasoning, all outputting to markdown files on disk. Then I handed those files to an AI engineering team I had built specifically for this, and told them to build.
The main functionality — entering API credentials for OpenAI, Anthropic, Grok; creating rooms; inviting personas; configuring their traits; conversations where the personas talked to me and to each other — worked on the first try. Small UI bugs. Sliding panes that weren’t as smooth as I wanted. But the core experience ran.
I sat there watching it and thought: I don’t need to write code anymore. I need to write specs.
This was November 2025.
A few months earlier I had written about something I called the Specification Pyramid: the idea that if you build your thinking top-down, from vision to architecture to features, the implementation almost writes itself. I thought it was an interesting framework. I wrote about it.
Then I used it to build Visionaire. Then I used Visionaire to spec the persona app. And when 38 AI personas started talking to each other on the first run, I stopped thinking of it as an interesting framework and started thinking of it as the only thing that matters in a world where AI can write code faster than you can read it.
In the months since, I’ve used the same Visionaire workflow to spec Rails apps, React and TypeScript apps, and Electron desktop apps. I tried it once with Tauri in Rust; the build worked, the idea didn’t, and I can’t remember what I was trying to make. The point is the spec never cared. Visionaire produces specs that don’t name a language or a framework. The same document built a Ruby web app and a Rust desktop app. Different stacks. Same source.
Code is the byproduct of a good spec.
At the speed AI can now produce code, the code stops being the moat. It’s produced too fast, in too many languages, by too many tools, for any particular implementation to hold value. The spec is what survives translation. The spec is what you keep.
The problem it solves
You have the idea. You can see it working. But between that vision and an engineer — human or AI — who can execute it, there is a gap that swallows time and produces wrong things efficiently.
Most AI tools jump that gap immediately. Paste your idea in, get a prototype out. Fine for simple tools. But for products with real customer journeys, onboarding flows, permission systems, multi-step processes, anything a user has to learn, the thinking has to happen before the building. Otherwise you’re generating code toward a target nobody clearly specified.
Visionaire is six phases that close that gap. Not by skipping it. By walking through it in the right order, fast.
How it works
Each phase produces a markdown document saved to disk. The interview, the decisions, the reasoning. A reference you can return to and ask: why did I decide that?
1. Idea. You explain the idea. Visionaire asks targeted questions, generated from your answers, to pull it out of your head. It does not add its own ideas unless you ask. An AI that starts generating before yours have finished forming is solving the wrong problem.
2. Product Brief. The conversation turns analytical. The problem, the users, why existing solutions fall short, what success looks like, what’s in and out of scope. Idea and Brief are deliberately separate phases because creative thinking and analytical thinking suppress each other. Doing both in one sitting produces worse output than doing them in two.
3. Market Analysis. A team of agents researches market size, competitors, and timing, and audits every assumption you surfaced — validating with sources, challenging with counter-evidence, or refuting outright. The phase that tells you whether you should build the thing, not just whether you can.
4. UI/UX. Before architecture, the customer’s experience. The end-to-end path through the system, the first moment that makes a new user feel the product works, accessibility and performance constraints. This phase lands before architecture because the journey doesn’t bend to the infrastructure.
5. Architecture. Opens with a feasibility pass over the UI/UX doc. Is anything technically infeasible, disproportionately expensive, or solvable a simpler way? Then technology stack, scale, security, integrations. Every decision in service of the customer journey it inherits.
6. Feature Specs. Visionaire takes over. It maps the customer journey to features, orders them by dependency, and generates one spec per feature. Filenames are numbered. Each spec carries its dependencies in frontmatter. Context-window-aware so whatever agent picks it up can execute cleanly.
One pattern runs through every phase but the first. Before any document is written, Visionaire produces an outline and asks you to approve it. This is the ambiguity check. The moment where assumptions surface while they’re still cheap. An ambiguity in phase two doesn’t stay in phase two. It shows up in phase four wearing a different hat, and by phase six it’s a load-bearing wall. The outline is where you catch it before that happens.
Get it
Visionaire v1 is available now as a Claude Code plugin. $297 at thevisionaire.app.
If you build complex products. If you run an AI engineering team and want to hand it specs good enough to actually build from. If you’ve spent days on a PRD and watched engineering pick it apart in the first review.
Visionaire is what changes that.
I wrote about the Specification Pyramid on this newsletter as an idea. I turned it into a tool. The tool built a desktop application with thirty-eight AI personas before lunch.
The spec was always the job. We just didn’t have a fast way to write a good one.
If you run Visionaire and don’t find the process valuable, email me within 30 days. I’ll refund it.