Just published: Endless possibilities in Umbraco with MCP, Cursor and one prompt. With just a single prompt in the Umbraco MCP server and Cursor, a brand-new Announcement Banner Block is built, published, and integrated into Umbraco, while generating all the view and style files you need. https://www.linkedin.com/posts/jeroen-breuer-3b135031_endless-possibilities-in-umbraco-with-mcp-activity-7361320073648566272-Whca

It's a cool example, but I can't help think that prompt took you longer to craft than it would have taken to actually implement it by hand. Of course, maybe an AI interface would be more accessible for some so there is that angle, but your main claims are around speed and fewer mistakes. The "fewer mistakes" is achieved by providing a thorough step by step prompt and leaving nothing for the AI to do but follow instructions. But IMO that has come at the cost of taking longer to describe than the overall task. I think the real power comes though when you cover a more complex task that would be much harder / time consuming for a user to do manually. Still cool mind 👍

ponders could you introduce another flow into the AI for "refinement" give it a small brief, but a larger success criteria, and let it self refine it's prompt if what it makes doesn't actual meet the final requirements?

Matt, I wonder if this is solved when you have a "prompt template" and you fill in the steps for what you need? Perhaps at that stage it is faster? Especially if you're making a large amount of blocks and you already have things written like this for completion criteria.

Yea maybe, but then equally at what stage is filling out a "prompt template" less user friendly? A good 75% of that prompt seemed specification of the new block + it's content, so you'd have to create that every time, but you'll have to check the template to see where you need to fill in what. Maybe if you could create some kind of reusable prompt that has a series of questions to ask you "What's the name of the block?" "What fields does the block need?" "What's the content for those properties?" "What block list does this need to be added to?" etc, then that could possibly be re-usable and somewhat friendly. But I've no idea what that pompt looks like 😄

@Nik that’s a common approach in reinforcement learning, but the challenge with it in this context is that you need a clear definition of success criteria.

In my demo the prompt was quite detailed to clearly show the full end-to-end flow. With a reusable prompt template it could be faster, but even if it takes the same amount of time, it’s still a good alternative. You stay in one environment, avoid all the CMS clicking, and have both content and code handled in one go.