I built tacit.works using a €20 a month AI subscription. I have never built a website before. The trick was role reversal. I didn't ask the bot questions. I gave it the role of PM and let it interview me.
The PRD chat
For all digital products I've worked on, there was a structured process of requirements gathering, triangulating needs and priorities between users, business and engineers, usually culminating in a Product Requirements Document (PRD). For complex projects this is heavy work. In my case, luckily, it was simple enough to be handled by a single chat thread.
I gave the AI the role of PM. Its job was to interview me and extract my ideas and requirements. It would use its training knowledge (product management best practices, web development) to ask me questions I'd never thought of, and give me options to choose from. Finally, it would use its powers of synthesis to produce a structured document out of all my unstructured inputs, one that I could hand directly to my "engineering team", waiting for me in another chat window.
The build
That was as simple as uploading the resulting 39-page document into Claude Code and watching in amazement as it got busy building the code. It occasionally asked for permission or for something to be clarified. My role was spectator and approver, ready to pull the plug if it did anything crazy. Often honestly out of my depth and not fully understanding the questions it was asking me. It only had access to a very strictly defined set of folders on my machine, so any damage could be contained.
Testing and changes
Then I had something built, but I wasn't really quite sure what it was. Time to test. A new chat thread in the same project let me keep the context the agents had already gathered.
My friendly assistant walked me through the process of setting up GitHub and a hosting provider, step by step. When I got lost, I uploaded the screen grabs and asked for help. Soon I was happily testing all the pages and compiling a list of bugs and change requests. I jotted down rough notes, then had my 'PM' work through them, suggest options, and produce another structured document. This time a change request, to upload back into the code builder.
This process I repeated several times as feedback came in from a kind set of early site testers. The site, complete with blog, newsletter and appointment booking, is now live.
My fractional PM agent
It's important not to overclaim here. I didn't automate my own PM. AI gave me the luxury of having a fractional, virtual version of one. A cost that would otherwise have been prohibitive.
My AI 'PM' was the brilliant intern we've read about so many times. Needs to be told everything in detail. Asks basic questions. Also an untiring and fast researcher.
What I had created here by default was my very own fractional PM agent. A smarter version of the classic prompting technique where you assign a specific role and workflow to the AI.
For something as basic as a freelancer's website, this is great. How far it stretches for more complex products, I don't yet know.
Things I wish I'd known
A few things that would have saved me time and expense:
- Use markdown files for the resulting PRD and change request documents. Much more efficient than Word or PDF.
- Both workflows burn through a lot of tokens, and I hit my limits quickly. That happens faster in the afternoons, as I learned the hard way. Try doing this type of work in the mornings, at least at the moment.
- Double-test everything. Even if the bot tells you it's all working perfectly, you will certainly spot things it dismissed.
- You can't avoid the terminal. For non-technical people this is intimidating, and even an anachronistic way of interacting with a computer. But you have an expert at your fingertips. Simply copy and paste the commands and responses back into the chat window when you need help.
- In the Claude Mac App. Create the documentation in the Chat, not in Cowork. It's easier and more reliable for the agent to hold the context of the thread in Chat. Cowork threads sometimes lost context, causing annoying rework. Chat threads can also be accessed easily from other devices on the go, which isn't the case for Cowork threads.