Coding With Claude Week 2: End of the Honeymoon?

This week, Claude helped me bring my son onboard to collaborate on our contract whist scoreboard and that worked very well.

Then I began a new project – a link management system with rich categorizing, tagging and drag and drop support. I wanted to test Clause on more complex graphical interactions (and make something that I will use every day.

But here I started to experience problems. I am not yet sure whether these are due to my prompting skills or Claude’s coding (or memory) skills …

AI Amnesia

I started out being frustrated because Claude didn’t seem to know the database schema that it had itself created. I was anticipating that it would remember what it had done in previous sessions, but it doesn’t. It starts fresh each time, even though it is working on the same repository.

So we created the .claude directory. Then I asked if it would think to update it:

Making Assumptions

So, I got into the habit of asking it to update the docs. But then I found that it wasn’t really thinking very hard about what it was doing and that it seemed to forget about context.md during the session

We started trying to get the basic behavior of dragging and dropping categories onto each other or into empty space working.

Incremental Guesses

And we went round and round in circles with fixing one bug causing another behavior to break until I wrote:

Claude came back with a suggestion for a new component, but it spelt it wrong:

I asked how I could help:

I haven’t set up local testing (yet) but I am providing full testing info.

Re-Thinking

When Claude started talking about overriding certain library behaviors for what I suspect is a solved problem, I suggested it revisit some of its assumptions:

And I had to emphasize that everything should be draggable:

We agreed on this simpler architecture, but then it made an error in a call to the database schema that it had created in a previous session (and which it was now supposed to consult before proposing a solution):

And again:

“Honesty” and “Complete Honesty”!

Now it needs to be “honest”:

And then “completely honest”:

But by then we had reached our usage limit for the session:

When we were able to restart, we put in place a debug log which seemed to help a lot and I got into the habit of reminding it to check its solution before implementing it.

And finally it was my turn to forget something. I didn’t update the database!

Running Out of Time

We burned through three complete sessions on Sunday and have used up 71% of my weekly allocation in three days! And we have something a basic skeleton of a solution that is not yet usable. Let’s see if we can continue in a more efficient manner!