At Rural & Co., we love long-form writing.

Long-form writing allows us to explore different ideas, to stretch, to take our time getting to the point. But sometimes, we just want to share something that doesn't need to be quite as expansive: a singular idea, a thought, or an update about a topic we find interesting. That's where Field Notes come in.

Or, think of it this way: if Parcel Post is a long, handwritten letter from an old friend, then Field Notes are unexpected postcards from someone you know (and hopefully, like).

In this issue, I'm going to talk about how I finally let go and learned to love Claude.

So, in this article, I am going to intentionally talk about my experience using Claude Cowork (which is a pretty good name, by the way). I will leave the technical walkthroughs to others, but I am more interested in the user experience of, and the design philosophies behind AI interaction models, and the different approaches and different interfaces that people are trying in order to facilitate adoption of AI-first workflows.

First a bit of background. I've been a ChatGPT Pro user for a while now, and I love the way that it It utilizes memory as a way to learn, and I also generally like the way ChatGPT handles projects, and I appreciate the ease of creating custom GPTs for dedicated work. It's served me well, and generally speaking, I feel like it has continued to improve with use and time.

Rural and Co. runs on the professional version of Google Workspace, and I was pretty excited to potentially use Gemini as my LLM of choice. But something has happened with the release of their latest model; Gemini feels frustratingly limited in its ability to think about complex issues, and instead seems to really only focus on summarizing content in my workspace, capturing meeting notes, and helping with administrative tasks within the Google ecosystem.

That leads me to Claude. It seems very capable and I appreciate many of the intentional design choices and overall brand direction: it feels very calm and organic. Even the off-white background and the salmon-colored accents are thoughtful touches. It feels very relaxed. And I don't think they get enough credit for their deliberate microcopy choices ("Coffee and Claude time?") and the overall voice and tone of Claude itself.

For knowledge workers, Claude Code is seen as the gold standard, and rightfully so as it does a great job of calmly and competently handling almost any coding tasks I can throw at it. It also does a great job of showing its work (more on that later). However, I only use the web interface and I never bothered with the terminal; I'm a passable developer and I have lots of experience using the terminal, but as capable as Claude Code is, positioning the tool specifically for developers in a development environment feels weirdly niche and "pre-AI" in a post-AI world where the workforce is evolving from role-based identities (e.g., "I'm a Developer") to skills-based identities (e.g., "As someone who primarily designs web experiences, I also build functional, interactive prototypes using such-and-such tech stack"). The MCP ("Model Context Protocol") open standard also shows real promise, allowing AI models like Claude to connect to external tools and data and do stuff. And this is where things get interesting.

Anthropic, the company behind Claude, noticed a pattern: developers started using Claude Code for everything except code (vacation research, photo recovery, slide decks). Essentially, their best agent was being used by people who had little interest in using a terminal as a way to interact with the LLM (like me, remember?). So Anthropic built Cowork in two weeks using Claude Code itself, with the thesis that agentic capability shouldn't require technical fluency. At first, Cowork was only available for Max users, but once they opened it up to Pro users (that's the plan I use), then I finally decided to kick the tires.

I usually capture tasks using a variety of tools (Todoist, Google Tasks, Notion — don't judge me 😅), and in an attempt to have One Task Management System To Rule Them All, I've been migrating to Linear as my project and issue tracking system of choice. As an initial Cowork experiment, I uploaded recent/active project briefs and meeting notes to Claude, then asked Cowork to migrate my scattered constellation of tasks to Linear, organize them, then prioritize them. Oh, I forgot to mention an important part of the setup: in order for Cowork to do its thing, I had to explicitly grant folder access and network access. Everything requires explicit user permissions in order for the agentic magic to happen.

So: I assigned Cowork the task. What transpired was my very first "A-ha" moment as I saw Cowork open browser tabs, scroll through sites, set up projects/tasks in Linear, and describe what it was doing during every step of the process. At the end of the process, I had all of my tasks added to Linear, tagged and organized into projects with associated dates, AND aligned to my 30-60-90 day roadmap for Rural & Co., which I had uploaded to Cowork prior.

It was, frankly, pretty amazing.

It was the first time that I could actually see the promise and power of an agentic workflow in practice. Even if right now it's a little clunky and requires some patience, it is clearly an early, successful example of how we will work and interact with our tools and data in the very near future.

Finally, a few notes about the UI and the interaction model. Claude now has a 3-tab interface: Chat, Cowork, and Code. Each mode has a pretty clear job, and Cowork strips away the "conversational theater" and focuses on execution. In other words, instead of asking Claude to do something, you describe the outcome, grant access, and watch what happens. At my previous agency, I would talk to my teams about "outcomes over outputs" constantly; it's really interesting to see that philosophy put directly into practice as a model for work in the AI era.

Also, I like the transparency of the interaction loop. It's an expansion of the "show your work" approach that's becoming a standard in AI interactions (specifically, points 2 and 3 below):

Inevitably, other LLMs will also become more capable in their agentic abilities. However, I am going to go out on a limb and say that the tool or service that will "win" actually won't be the most powerful or even the most feature-rich. It will be the one that simply does things competently and provides the best user experience.

I recently reposted a link from an article by Microsoft researcher Suff Syed in which he provides the following observation:

For agentic workflows, replace the word "vision" with "experience" in the last sentence of that statement, and it sums up my thoughts on the future of agentic workflows quite nicely. In a world where nearly everyone can execute with the right tools, experience is upstream of everything, and execution follows.