AI codes fast. Until it’s time to ship, or iterate on a complex product.
Then nobody knows what the AI did, what you did, or how or why certain decisions were made.
A reliable product development methodology does not survive code thrown together at inference speed.
More than a year ago I posted a small idea on GitHub about Cursor: instead of improvising prompts, start from a specification like PRDs, user stories, acceptance criteria, and make that the basis for the code. Spec-driven.
The repo reached 340 stars and stayed there.
I didn’t abandon it because I stopped believing in the idea. I left it because the idea was only the beginning. Since then I’ve kept working in the shadows, many nights, many weekends, time taken away from my son and my wife.
Today „spec-driven“ is fashionable. When I started, it was an intuition. That at least tells me the path made sense.
I’ve published PaellaDoc: the way I’ve been developing software for years, now turned into an app and taken as far as it needed to go.
The thesis is simple: working with AI doesn’t have to be a black box, nor does it have to turn developers into product managers, or PMs into developers. Everyone has their place in this shared domain that is agentic development.
If you combine product discipline with development discipline, you can orchestrate Claude Code, Codex and Gemini on the same repository with the same rigor as without AI.
The flow adapts to how you work, not the other way around:
- If you already have a live project, PaellaDoc does reverse intake: it extracts the User Stories, Acceptance Criteria, epics and PRDs that are already implicit in your code and your history.
- If you have your own PRDs, you import them and work from them to generate aligned code.
- If you come with your methodology (Jobs to be Done, Shape Up, yours), you plug it in as a plugin for the open SDK.
- If you’re starting from scratch, you define what you want and PaellaDoc helps you structure it with product discipline.
From there:
- PaellaDoc distributes the work between Claude Code, Codex and Gemini with the specialized context of each role. Also local AI like Qwen and others, but honestly I’ve added that and still haven’t tried it.
- Playwright Golden Gate verifies that the app actually works before closing the task (in alpha).
- You review something finished, with its ACs being met, not an 80-message conversation.
Some important details:
- For now, only Mac Apple Silicon. The rest will come if it makes sense.
- 100% local on your machine. You provide your API keys or connect your Claude Max or ChatGPT account; nothing goes through my servers, so it burns through tokens like you wouldn’t believe. Don’t try it without a MAX or expensive Codex account.
- No sign-ups, no account, no strings attached.
- Telemetry yes, but only anonymous usage and error events to improve the product, never your code, prompts, files or paths.
Enabled by default, it can be turned off with a toggle in Settings (and on first launch the app offers to disable it with a click before sending anything). It’s documented down to the last detail on my website. - Free for personal use, and it will stay that way. For commercial use in companies there is a license, which will allow me to keep working on this without the personal version changing. I haven’t found a fairer model.
Important: don’t judge me as a salesperson, because I’m not one. I’m on my own. It’s free, it’s usable, and it’s far from perfect. There will be bugs, edge cases I didn’t see, debatable design decisions and, above all, things that STILL DON’T WORK.
It is what it is, and I prefer to show it this way, in ALPHA, than wait until it looks polished in a demo.
What I need most right now is criticism. Especially the harsh kind.
What you don’t understand, what feels off, what’s unnecessary, what’s missing, what you think is poorly thought out. Use this forum to discuss it seriously, every specific criticism eventually becomes an issue I prioritize openly.
My availability for this is the weekend, except for the odd spare moment. I ask for your patience.
And in the end this is what really matters to me:
Between the purists who say „AI should not write my code“ and the vibecoders who let AI write everything without looking, I think there is a middle ground. A new paradigm where you embrace the tool, but with the usual discipline.
I want to make my small contribution there. And even if in the end only I and four friends use it, that’s enough for me. It helps me every day since, a year ago, I decided to work spec-driven, and then introduced the system at Rankia. That’s why I’m sharing it.
If it helps you, repost it so it reaches more people. If it doesn’t help you, write to me why, seriously, that helps me more than a like.
P.S. if you try the tool and find something broken, tell me here. No filters. I won’t be offended, it helps me.