For the UX piece, giving it access to "see" it with Playwright helps tremendously. It fixes things I didn't know were wrong. But then when I look at the new UX, it is clearly better. It's wild.

Half the time when I'm finishing up on some UI (that I am terrible at) I'll ask Claude to "make it better".

"Sure thing boss"

Basically every software engineer in my life is obsessed with Claude Code and loves it and say it feels like an absolute step change, and these are people who've been coding for 30 years.

It was right. I gave up and removed it for good.

I created a pipeline for something similar. I had probably had an easier time with some good metadata to work with, but it was grouping reasonably well at the end.

github.com/schuettc/ple...

I’ve got one where it keeps suggesting I just remove a component because it intermittently fails and can’t really be worked around. I keep arguing with it trying to find a workaround. I’m pretty sure it’s right and I do just need to remove it :(

I think one of the most important aspects of using things like Claude Code is finding something that works for you. So far, this is working well for me. It will likely change in a few weeks when something new drops.

Besides the technical aspect of not polluting context, I've found this helps me focus on shipping meaningful improvements rather than getting lost in a maze of half finished changes. It's all set up and ready to work with multiple developers working on the same project at the same time.

It lets me easily capture bug fixes, new features, enhancements, tech debt, etc. all without disrupting my flow. Then, when I'm ready to implement, all of my notes and thoughts are ready to go.

One interesting aspect of this process was using github.com/microsoft/pl... to have Claude capture images for the documentation. I know it's dumb, but I always slack on getting good images as part of the documentation. Having Claude do it via playwright was great.

I haven't tested with all varieties of LLM provider so there might still be some rough edges, but wanted to push something out there rather than fiddling with it forever. Issues and PRs very welcome.

I think it ended up much better and more accurate. It might be a little too aggressive at creating collections now, but that should be pretty tunable now that I have a strong embeddings mechanism.

It started simple enough with a pretty basic single pass LLM request and was working ok. But, as is often the case, when the project was ~80% complete, I started to fiddle and adjust. The result was an almost complete rewrite of the LLM process.

youre-absolutely-right.md

And it's hard to teach them experience. But I imagine it's frustrating for someone without it to see someone else 5x when they are struggling to get anything meaningful out of it. So most of what I've been doing is trying to write SKILLs/commands/etc to help guide them.

They could... :)

That's how I learned too, but I don't know that many are doing that today. The whole landscape is really weird and changing so fast. I watch other people try to use Claude Code in horror. But it's hard to help them because it feels intuitive to me because I have more exp.

I've changed my workflow half a dozen times in the last few weeks. Each time improving on it. Hoping to write some of this down in blog/GitHub. Realized today that I created a harness for myself....

But that's not all. I'm also doing a session on A2A that will knock your socks off. Come see how to deploy agents and watch them work together completely dynamically.

registration.awsevents.com/flow/awseven...