This is super interesting. Would you be open to recording a Loom video where you walk through this a little bit? And I'm curious, what's the text app? Like, what does the data flow look like? Is this like a fully automated workflow, or is this something where somebody's driving Claude, let's say, with a bunch of connectors with read and write access to the different systems?

The Neptune acquisition is interesting. I had not been aware of that one. Persistent memory is clearly a huge challenge and core to the user experience. I'm a little bit skeptical of anything short of more frequent local weight updates, since I think that's the most general solution, but I don't know how far away that is. Could be two years, could be 20. Although maybe I'm wrong, but I did a little cursory research and it seems like monitoring and observability tooling for training runs, or am I getting that wrong? That would make a lot of sense as an acquisition - a core piece of infrastructure basically for them.

I agree that the role is shifting, and I think thinking like an architect applies pretty broadly. I guess I'll share two ways that I've fundamentally changed my workflow. One is I try to spend ample time with the end user, having them explain in excruciating detail their existing workflow goals, the different systems they're interacting with, and so on, and record it all in Fireflies, meeting recordings and transcripts. Usually one to two hours worth of that content is sufficient to then pipe into a Claude Code session, along with some additional dictated context if I have ideas or additional context that wasn't documented in the transcript that would be useful to someone who's doing an actual build. And then I'll let Claude Code rip and work with it until I'm satisfied. The other side of it is thinking about using AI for the purposes of QA and QC. Automated test writing is, of course, great. And using plain language to say, I want higher levels of test coverage, or what can I do to make this more robust or easier for someone to clone the repository and contribute to themselves. I've started adding lots of bots to do automated pull request reviews. And so if I'm ready to submit a PR, I'll tell Claude Code to, after it's submitted the PR, wait for 10 minutes or however long it takes for those bots to do the reviews, make comments, have it review the comments and prioritize additional changes that it can make. And then overall, I find that that improves quality in a way that makes the project more robust. Getting really clear on the tacit knowledge that you're trying to unstick from subject matter experts and incorporate into a new piece of technology. And then the different success criteria that you'd want to validate it against seems to be where the surplus attention should flow. And that's hard for people who come from an engineering background because they're not necessarily used to spending that much time with customers. I come from the opposite direction where most of my career has been in discovery, launching new products, process transformation, et cetera, with technology always as a means to that end. For an example, here's a project in an open source repository for a command line tool that allows recruiters using the Greenhouse applicant tracking system to massively accelerate applicant review. I built it using this methodology I just described.

One thing I'll add to this is somebody who's been doing this now for three years is that step zero before you can get into this kind of messaging is to create compelling facts, as I like to put it, meaning actually create value and then talk about those stories. When it comes to good writing, the details are what makes something come alive. And instead of speculating, helping out someone, even for free if you're just getting started, and then recording the process end to end with Fireflies, capturing digital exhaust and email interactions and so on is a great way to go. That's how I draft all of my case studies and some of my social posts. I'll take transcripts where good outcomes were achieved or I learned something surprising when it came to applying the technology, or for multi-week or even multi-month engagements, getting an overall dump of Fireflies transcripts, emails via MailMeteor, and so on, and then piping them into Gemini 3 Pro, which has a super long context window and can generally handle that volume of data and asking it to draft compelling stories. That raw material becomes an asset that you can repurpose and point at particular personas that you're trying to reach. Don't skip step zero!