Fishing Scout

About a month ago, I decided to start a project where I would use Claude as an AI app builder to build a new web app (almost) every day. Most days, these apps were completely frivolous, and I simply deleted the project and moved on. In any case, it has been a good exercise to stretch muscles thinking about software, digital products, and solving many of my own personal bugbears.

Our oldest child is still a toddler, but has been expressing an interest in fishing for the past week or so. They are too young to really be able to understand what they need to do, but last summer I came up with a method for them to be able to fish by eliminating one of the variables, setting the hook. Using circle hooks meant we missed far more fish than we otherwise would have, but it meant all I really had to do was keep them distracted until a fish had hooked itself and then redirect them to reel the fish in. This type of fishing is typically meant for rivers, so yesterday when they mentioned it again, I decided to prod Claude on if there were any local lakes where I could attempt to target small channel cats with a toddler. I then forgot about it until this morning.

When I came back to the chat and saw the response, I knew it was not helpful. The suggested techniques were correct, but the lakes it suggested absolutely do not have a meaningful quantity of catfish, and the suggestion to fish the Minnesota River is precisely what I am trying to avoid with a toddler. When I pointed this out and requested lakes that have a higher CPUE based on recent survey data, Claude said this was out of scope of what it would be able to answer and suggested that I look on the MN DNR website for more information.

So I said fine, let’s build the app. I defined my architecture, required features for an MVP, provided links and samples of the JSONs from the MN DNR APIs. I provided some notes I had on the data within the JSONs after looking at the responses, and how I thought we could work past issues.

In less than 15 minutes I had an MVP, and after consuming all of my credits for the session, I had in place pretty much anything I could want the app to do. Rather than eating into extra credits, I fixed some bugs myself and executed a few curl requests as a workaround for the one bug that I will wait to have Claude resolve in a future session.

The thing is, now that this app is available to me, theoretically the output or database the app can create would be useable by Claude to answer my initial question.

What does all this mean? I am not really sure yet how everything will shake out, but if I had a crystal ball, I would guess that LLMs and AI agents will start to recognize when they are running past the limits of their training data and tools, and from the prompt itself start to create tools necessary to give them what they need and not require the direction from a prompter to do so.