Today I crossed the line from using Cursor as a product thinking partner to having it build my actual product (not a prototype, the real thing).
(I’m building a desktop utility that is like a screenshot tool but instead of images, it OCRs the text and sends it straight to your AI’s context)
Previously: I’ve been using Cursor to hold all my research, scoping, and product strategy in markdown files. Just text documents AI can read.
This morning:
1️⃣ I pulled up a lightweight opportunity assessment for an MVP I’d been thinking about (I can’t emphasize how much time, research, thinking, and manual offline prototyping this took. Recognizing and scoping this was the biggest step of any of these.)
2️⃣ I asked Cursor to turn it into a detailed SPEC.MD using a prompt I’ve refined across multiple side projects
3️⃣ It interviewed me. This part felt oddly similar to my best engineering partnerships. It pulled out decisions, surfaced edge cases, and challenged my assumptions
4️⃣ In the same thread, I asked it to create an implementation PLAN.MD following “something simple working soon, then iterate”
5️⃣ I had it write an AGENTS.MD file (instructions for how to work in this specific codebase-combining general best practices with decisions unique to this project)
6️⃣ As I write this, it’s building
The counterintuitive part: I kept expecting to shift into “engineer mode.” Instead, I’m making the same product decisions I’d make with a senior engineer, just faster and more iteratively.
My takeaways so far from the experience:
- As strong as Cursor was technically, can’t be autonomous completely, still need my help as a human to close some feedback loops (i bet this will go away rapidly)
- I thought today would be all coding. Turns out, it was still a ton of wearing the PM hat and having product-eng type negotiations/collaborations.
- This is so much fun and flow
And so can you!
After I published the above, so many of you replied to ask for the prompts, that I decided to tidy them up and publish on Github (cloud folders for sharing and collaborating on code… like Google Drive with more process). You can find my prototyping prompts here: https://github.com/talsraviv/from-thinking-to-coding. You can take that URL, give it to Cursor or Claude Code or Codex, and tell it to download the files (nerdspeak: “clone”). That way, your coding agent has the prompts as local files on your computer, and can read the prompts directly. Here’s a video walkthrough of how to use them in sequence:
I vouch for these as far as getting a highly functional working prototype live. If your goal is building a software business with customers in production, there’s a lot more involved. Colin Matthews has an excellent new course to those ends, and he and Peter Yang just released an excellent screenshare tutorial (not sponsored, just a huge fan).