In Report 1, I mentioned how I was able to get an app working locally on my laptop within several hours. To recap, this is an app to collect news items exposed through feeds and create a searchable archive that can be used in research. In my next iteration of the app, I changed the database (MySQL to MariaDB) and deployed it on a server. The app is now available here (URL is temporary, trying to think of something besides some form of “newsarchive.com”, but that may be where I end up). I also had Claude create some user documentation which is linked at the top of the app.

For the most part, Claude provided the instructions for how to deploy the app and connect to the database, and the instructions were mostly correct. If the instruction failed, I pasted the error into the Claude terminal, and usually Claude came up with a correction. The biggest problem was trying to get a cron job set up to check feeds once per hour – the provided job entry was not working. After some web searches and Gemini checks/fixes, Claude figured out that part of the command left out a dash on a switch option – ugh! I hope that has been fixed now…

Issues that I have seen have been due to the input feeds (NYTimes feeds have some duplicate stories across their feeds). Also, since I have only been collecting data starting today, the archive is someone small, and only has feeds from two sources so far (NYTimes and ABC News). It was disappointing to find that CNN does not have feeds any more…. I am planning to work up to adding at least one new source a day, but will need to find feeds!

All feedback is welcome – give it a try!

In a previous post, I mentioned that my first AI project was to develop an application to create a archive of US newspapers based on RSS feeds. I initially used Claude Code to review a list of US newspapers from Wikipedia to search for websites and RSS feeds. As I posted earlier, this took several sessions to complete this task, and I had doubts that I would be able to create an app without hitting my Claude Pro token limit.

My next step was to create a Markdown document describing the app for Claude to use as starting point. I looked at several example product requirement documents (PRDs) and frameworks for developing apps using AI (Get Shit Done, Superpowers) to find a way to get started. Eventually, I decided to follow Peter Yang’s advice on developing apps:

1. Don’t know what to build? Just tell AI your problems. Ask it: “What are some simple apps that we can build to help me take time back?”
2. Create a robust plan first. Ask it to “Create a plan with 3 clear milestones” to avoid mistakes during coding. Spend half your time planning with AI.
3. You’re the manager, AI does the work. Give it feedback and direction like you would a human employee. Be patient with it 

I created a initial plan based on some PRDs I found in Peter Yang’s Github repo. I then asked Claude to review the plan and see if there was anything missing or not clear. Claude responded with a set of 12 things to change and 4 questions before starting any build work – wow! I accepted Claude’s suggestions, provided answers to the 4 questions, and Claude updated the plan (see here).

In the updated plan, four milestones were identified. Claude was prompting me to say “start building milestone 1”, so I did it. Claude then took 10-15 minutes to create code and tests for the Milestone 1 content. I decided to have Claude confirm steps with me as it went, so I could see issues firsthand (and there were issues, but Claude figured them out). If I had “let it run”, it might have finished quicker. During the Milestone 1 work, Claude was able to figure out that I had Wampserver installed on my local machine (which has MySQL, a specified part of the app tech stack), and used that to create the MySQL database for the app. When Milestone 1 development was complete, I did some operations with the app, then told Claude to create a handoff document (handoff.md) in case I hit a token limit during the session.

I then repeated this cycle for Milestone 2, 3, and 4. The handoff document was updated after each milestone. It turned out that I did not encounter any token limit. When the development was complete, Claude then added all of the code/files to a Git repo in the working directory. Here are some screenshots:

App admin screen

App search screen

To me, this was very impressive. It took less than 2 hours to get to this point – a fully functional application with 79 passing tests – amazing! I will post more when I have a version of the app running on a server.

PS – I have created a new category for these posts (AI for Smart Old Farts), if you want to just read these posts.