AI Doesn't Have to Be a Solo Sport. Until It Is.

I want to start by adding that this is partially in response to something someone else recently wrote, I think. I haven't been able to find the article, however, it was referencing that AI is a solo sport, or at least I think it was. That premise, or my (mis)interpretation of it as intending that AI is a tool you are better off using alone, and is one that isn't good for / built for collaboration, has been rattling around in my brain since I saw it (and memory is a fickle thing; I'm not claiming to remember it's whole intent). What's following is what I have been thinking about, while working through a project that felt largely solo at points.

When I say "solo" in this article, I don't mean truly alone. I mean working with AI as the only other presence, without other humans in the loop. That's a real mode, and it's worth naming so the rest of this lands clearly.

Where the Team Part Starts

There was a recent inspiration for an MVP that had a quick turnaround. I'm certain the drill isn't entirely unfamiliar to a lot of folks: someone said something that identified a specific need and felt like it had some urgency implied to it.

It started at the end of the day with what I think of as the proper opening move: brainstorming with a couple of colleagues in different contexts to work out what a possible solution needed to look like.

In the first pass, there were more questions than answers, a lot of "could it, should it, why would it" type of things, and I was busy with sticky notes sprawling across doctor-signature-quality notes and awful sketches. I'm going to go ahead and call this the Discovery phase. We were trying to land on the actual problem worth solving, not just the one I'd assumed was in front of me.

Then a second pass with different colleagues, this one much closer to requirements gathering. We worked through existing infrastructure and recent delivery approaches, identified what data already existed and what needed to be synthetically generated for the MVP, sketched a definition of done, and made a handful of other less-important, yet critical decisions.

Those two upfront passes were rapid for me, in that they possibly took less than an hour. There was a shorthand that existed with the colleagues that comes from time spent working together, however, they carry more upfront weight than people realize. In this case, I wasn't really thinking of it as collaboration; I was getting to clarity as quickly as possible in an abstract way that I know how to do.

What struck me was how easy it would have been to sit down with a prompt and start typing in an unstructured and wandering fashion in this AI-assisted world. It's kind of like jumping right in to Figma or any other "drawing" tool instead of spending time with pencil and paper sketching through the thinking so that tool time is less, and better spent.

Those pencil and paper moments, whatever form they take, are where the work gets focused before any building happens.

Where I Went Farther Than I Expected

After working through fairly high-level Discovery and Requirements sessions, I then did the thing that I think gets mistaken for collaboration more than it is: I checked in with ClaudeCode to gut-check my development and infrastructure environment. Folks, I'm a designer, and I've been tinkering with tech since my TRS-80 on a 300-baud modem dial-up was the escape hatch from small-town life, and I'm here to admit that I'm scratching the surface on what goes on under the hood in my own environment, at least on a lot of things.

(Don't get me wrong; I've got some solid security, monitoring, and change control systems in place, however, they don't stay locked in my memory as well as the UX work that I grew up in industry with. I frequently need a refresher, and that's exactly why my heavily-documented system and agents exist.)

I needed to look at previously delivered systems that were similar, yet also different enough to be useful comparisons. Claude Code & Nucleus (my Qwen-powered internal agent; think OpenClaw without all the drama and/or snake oil) know the environment well, and I'm meticulous about documentation to make these types of situations happen smoothly, and accurately, so I can make informed decisions, and quickly. I don't see this as collaboration. It's engaging my confident interns to report back with qualified facts and evidence.

Once I understood the tech infrastructure and had made some development decisions:

• Flask app sitting on Postgres, specifically Postgres with vector search, meaning the database can find passages by how close they are to the question, not just by matching words.
• Embeddings on one HomeLab GPU, the language model on another, which is just how my internal HomeLab is set up; a lot of workload is separated out and points of failure are limited as much as possible.
• Hybrid retrieval; I can filter on structured stuff like document type, entity, or a contract vehicle, AND rank by meaning.
• Word and PDF documents accessible end-to-end, so users can find what they need with full context in a single click.

I kicked off deep research sessions to identify the right structures for the synthetic data I was going to need to get this off the ground. That allowed me to not only identify the document types I would need, it also allowed me to fill them in with synthetic data, and create the documents. That led to data ingest, chunking, tagging, and readiness for use in this RAG app.

Then I pulled from the requirements to get tasks aligned and understand what could be automated with minimal oversight. Over the course of an evening, I had most of the context I needed to validate data, map and re-map a data ingest process, and get ready to wire together a local LLM-to-RAG querying engine.

That meant I was ready to start building a front end and testing how my prompting skills would deliver synthetic content. Within 24 hours of scattered moments of focus, I had a real MVP. (Yes, that's easy to see as one of those AI flex things. I'm not trying to wow you with speed; it's not the point of this article. It is simply the reality of the situation.) I had delivered something that ran on real (albeit synthetic) data and could get real (and this means absolutely real) feedback, beyond what a wireframe or a clickable Figma prototype could ever offer.

I want to be transparent about the caveat: I have AI infrastructure that's been documented and tuned for the way I work. A lot of people don't have that yet, which means a lot of people would not move this fast on this kind of stack work. I paid the setup cost for that infrastructure earlier so I could move quickly later, and that's worth calling out.

This was the part of the work where it felt the most solo. I went farther "alone" with Claude (and Qwen) than I could have ever gone before. I relied upon my confident interns for infrastructure readiness knowledge, and then I "vibe-coded" (which, come on, is a bit bullshit; I planned and orchestrated this) my way through a remote-access dungeon and came out the other side with something that worked.

Where the Team Part Came Back

Once I had the initial prototype loaded with our synthetic data, I shared it out to a more technical colleague to understand how a development team might extend or rebuild this in production. The choices I'd made, from Postgres because I like it, to certain stack decisions because they were good-enough placeholders, were now visible, defensible, and ready for critique. The artifact did most of the explaining, however, no credible designer ships off a link without an explainer, right? Right.

The colleague's final statement was "I can align with that." And yes, I asked, to be certain. I wanted to know this meant it was a feasible way of working, and that the outcome could be replicated on systems not exactly like my own.

That kind of alignment matters. It came after looking at something real, and the conversation we had was specific because the artifact was specific. That's the show-and-tell moment, and it lands differently when the thing actually runs.

The Rhythm

When I trace the shape of this, and similar non-Russ-specific projects, they tend to follow:

Teamwork (Discovery) → Teamwork (Divergent, Convergent Thinking) → Solo + AI (System Scoping & Planning) → Solo + AI (Building) → Teamwork (Critique & Review) → Solo + AI (Iteration/Revision) → Teamwork (Delivery, Re-Discovery)

Two team passes happen upfront. The first is a quick-turn Discovery to get focused on solving the right problem. The second is more of the Design Thinking variety, where the solution starts to become more apparent.

Then two Solo + AI phases happen. The first is infrastructure and systems scoping and planning, where I get clear on what's needed to deliver the outcome and what's already in place, plus any compromises or new considerations. The second is the build phase, where everything that was planned gets orchestrated across a variety of agents. Data gets defined and structured, organized and created, ingested and implemented so it all lives under one roof.

This sets up more teamwork; MVP review and critique, followed by another Solo + AI round of iteration and revision, which can be cyclical. One more round may not be enough when you're coming in hot like this. And then, of course, more teamwork with MVP delivery. This, too, can be cyclical; it may turn into more iterative passes, or it may turn into another pass at discovery to ensure the right problem is still being solved and that the outcome aligns with the expectation.

Before someone says it: yes, this skips usability testing. At the MVP-validation stage, the question is whether the hypothesis is even worth putting in front of a user, and once it earns a green light, users enter the chat. Something often needs to exist before it can be good, and it definitely needs to be good before production.

The solo beats in there are real solo beats. I was working solo; I work through processes I've used since I was just a baby UXer, bringing the thinking, sketching, and defining what done looks like, and then deciding what to bring back to the team. Those parts are mine, and AI doesn't write them for me. AI is there, and AI will love to tell me how great and load-bearing my thinking is, however, I feel pretty confident in my ability to understand this current model's ability to placate.

Why It Feels Solo When It Isn't

Maybe I've got that backward. Maybe it doesn't feel solo when it is, because AI can seem like it's a real companion, and it certainly can do real, guided work.

However, here's the thing I keep coming back to: the reason any of this can feel like a solo sport is that I can now go so much farther before I need, or maybe it's want, to check in with others. The cost to getting "just one more thing" completed is low, and not at all low at the same time (ask anyone trying to tweak a layout about what happened to those 5 minutes they promised it would take, from over an hour ago).

The check-ins on incomplete work that used to happen at every stage can now get collapsed. Maybe that's ego, or maybe it's the benefit of new capability. I'm certainly guilty of the ego part and showing off my completeness of thoughtful thinking at an MVP, just because I can. That said, I no longer schedule a meeting before I've started working. I no longer share a half-formed wireframe just to keep people in the loop, though I still happily "share before I'm ready" with anyone available or willing to sit through a rapid-fire review.

In the old way (you know, 2023-ish), I'd have checked in more frequently and with less to show across more time. The New Way™ eliminates a lot of that, without removing that this is still an MVP. The end state is the same kind of artifact, with much more systemic thinking and planning baked in by the time anyone else sees it.

There's also the matter of availability, and it might be the biggest piece. AI is there when I'm there. I can be awake on my iPad at 10:30 on a weeknight, hit with an idea, and act on it right then. I don't need a colleague to be awake or available in order to run "hello world" on an idea when it strikes. A human collaborator has a calendar, a life, and perhaps more important than a well-defined supplements stack: a sleep schedule. AI has none of that, so it's available exactly when I am (token availability pending, of course). That's part of why the work feels solo: I'm never waiting on anyone. And it's part of why AI feels like a collaborator: it's always there when I want to think out loud.

There's a second piece here that's worth calling out, because it's where the "we" language gets misleading. Human collaboration has divergence built in. Your collaborators can identify blind spots you haven't named, or say "wait, you should consider doing this other thing" without being asked. AI doesn't usually seem to do that, at least not yet. AI follows the direction I set, faithfully, and it executes on the threads I pull. It doesn't pull threads I haven't pointed at. Oh, the time it would save me if it did...

A recent example: I have an AI agent called Nucleus running on a Mac mini, tapped into a Qwen model on a different machine. I'd been complaining about Nucleus being slow for weeks (so much for all the Open Claw hype). AI didn't tell me "what if we ran a smaller local model on the Mac mini itself for quick inference, and routed bigger queries outbound?" I had to pull that thread on my own, and once I made the connection to a previous way of working locally, the answer was obvious and the work to make it happen was trivial. However, surfacing the idea was on me.

That's the difference between human collaboration and AI collaboration. Both are real, both are messy in their own ways, and only one of them can pull me off course toward something better, with intention.

When I say "we," and I do say "we," I mean my AI counterpart and me. That's true, yet it's a different "we" from the "we" I use with another human, and collapsing the two of them into one obscures more than it reveals.

It's Just Me, Myself, and AI

The hype gets this wrong; none of this is a revolution.

This might be how we work now, or how we work next, because we can, and because we can shape it to fit how we need to. Working this way is a bit like a favorite pair of jeans or a pair of high-quality shoes; it takes some breaking in before it takes the right shape. Off the shelf, it's just a tool, and it'll stay that way until you take the time to break it in. Broken in, it's something else.

The demand to explain and defend design decisions didn't go anywhere, and no magical skills.md files are going to change that. It shifts; it is shifting now, and will likely continue to shift more and more as we all make our way through this.

All of this "we" business is the part that makes it easy to conflate what happens when you're going it alone with AI in the sidecar. It can feel an awful lot like you've got someone along for the ride, and it can feel an awful lot like you've got a real-life collaborator.

And you kind of do. Until you realize that you also don't.