Experience Building with AI - Jan 2026 Checkpoint

Writing a quick post here on my experience currently building with ai as much from my own reflection as any thing (stream of consciousness, may refine later).

I've been using Claude code a lot. I like to use Conductor with Claude code and then I'll jump around using cursor and then Claude desktop and open ai for simpler things occasionally. I found there's a ton that I can do and can do really quickly that I just would never have done before. A lot of desk research and time consuming analysis is incredibly quick to do now so that's definitely been a cool part of the experience Paradoxically I found that as others have noted it still takes a lot of planning and coordination especially for things that aren't software and even for software to build something that's really durable nails what you're trying to do and works really well so I'm simultaneously impressed with what can be done and haven't found it to be a massive unlock that I can really unleash on its own just yet.

I'm having some of the common challenges that I see others building have namely it would be great if long running agents were cloud based, not local, is a big one getting good end-to-end environments including testing is another big one because that can certainly be done but is not trivial to do in different environments. And then there's also a lot of in terms of using it to really create output and share output with others getting a good form factor. Creating especially PowerPoints and PDFs remains pretty tough so it's definitely a barrier to accelerating my workflows.

I think that I could likely use more inference more effectively by spending more time setting up my my agents and architectures but always tough to know when to take gains versus continue to iterate on the workflow setups.

The thing I'd like to try next is further utilizing the compound engineering framework from Every to try to document more of my objectives and preferences to allow for longer running agents.