The missing middle of AI context

First issue of What Stuck. Once a month I write down what actually stayed with me, which turns out to be a lot less than what seemed interesting at the time. For your benefit.

What I studied

My own setup. Most of this month went into Claude Code — skills, agents, hooks — and then into the workflow wrapped around them.

The AI speed boost I kind of take for granted now. I’m getting days’ worth of results out of hours of work. What I didn’t expect is how much more effective I am, rather than just faster. AI turned out to be a great sounding board, and better at coordinating my work than I am.

That only holds if the setup underneath it is good, which is where the time actually went. I have skills and agents that flesh out an idea, turn it into tasks carrying their own context, problem description, acceptance criteria, and definition of done, and keep the board groomed as things move.

Another set picks what to work on next against my current focus and deadlines, plans the implementation, reviews the changes critically, and bundles the result into atomic commits with full traceability.

Over all of it sit higher-order skills that shepherd a task through a different process depending on its size, complexity, and risk.

I’m currently writing all of this up as a proper article, but it isn’t finished just yet. Until it is, the raw version is in my dotfiles — the skills and agents themselves, only without the explanation.

A tool that earned its place

I usually have several Claude Code instances going at once; one to three per project. One does the main development work, one plans what comes next, and one digs through the code to help me make sense of it.

I ran that in tmux for a while. Claude would ping me when it finished its work, and the associated tab would highlight, which worked, but the whole thing was rather messy.

I know it might not be healthy, but I want to know that Claude is working for me around the clock. Anything less feels like a missed opportunity.

Herdr is what I moved to. It does tmux’s job, built around orchestrating agents rather than shells: an overview of every project, a terminal for each task inside it, and the thing that sold me — one global view of every agent running on my machine, with its status.

I can see them all working, and the moment one finishes or needs an answer, I know, and I can go and give it one.

I was impressed enough to build something on top of it. Workbox is a Terraform setup that provisions a cloud VM (GCP for now) with Tailscale for security, and a CLI to drive the machine. It sleeps and wakes on a schedule, or by hand when I want it to.

It uses herdr’s remote feature, so I monitor and control the agents while they work on the cloud machine instead of mine.

Finally, I can close my laptop, and Claude keeps going.

A book worth your time

The 90 Day Brand Plan by Dain Walker — five stars!

Leaving my job and starting Seastone was, in large part, about taking more command of my own life. Really owning my personal brand and being deliberate about how I show up in the world is a big part of that.

From a business perspective, I think that my personal brand will have a huge impact on my success as a consultant. While I know I’m really good at what I do, that counts for very little if nobody else knows it.

So I dug into personal branding. Five books later, this is my favourite.

Walker’s main message: brand the person you’re becoming. Defining your ideal persona and then acting accordingly turns you into them.

What I like is how that enables a beginner’s mindset. You don’t have to be an expert at everything, and showing that you’re still working at it will probably do more for your brand than pretending you already know it all.

His advice on content is blunt: don’t plan, don’t polish, just post it and learn. I’m not fully there. This newsletter has been polished more than he’d approve of, but it’s actually the reason it exists at all.

My full review has the rest.

From the industry

Anyone who has worked with AI for a while knows that the context you hand it decides what comes back. There’s a whole discipline formed around this, context engineering, and it keeps arriving at the same finding: just piling on more material stops helping quite early. What matters is whether the context is curated.

My own evidence is the notes archive I’ve kept since reading How to Take Smart Notes — reading notes, routines and checklists, meeting notes, reference material, shower thoughts, scratchpads. Years of it. Armed with all of that, my “sounding board agents” are doing an eerily powerful job of pushing back on my ideas and reframing things I thought I had settled.

What’s odd is that the industry is pushing on this from two ends at once, and neither of them meets in the middle.

Tools like Claude, ChatGPT, and Goldfish build a memory for the individual user. That works well. It also can’t be shared, so the better your context gets, the more it’s stuck with you.

Services like Atlan, Collibra and Glean come at it from the other end, gathering organizational knowledge into one shared context. That works well too, but it’s far too coarse to be really efficient for the individual.

What’s missing is everything in between: the tooling, principles and workflows for granular control over who exchanges which context with whom. An AI context mesh, almost, where every department, team, project and person gets what they need at their level and nothing beyond it, because the extra just confuses the agent.

We worked out how to do this for code a long time ago. We haven’t for context, yet.

Visiting AI Summit Barcelona this month made me more sure of it. The vendors either sell you the foundation to build it yourself: databases, storage, lookup. Or they skirt the problem by picking the individual or the organization end and stopping there.

Meanwhile, every attendee I spoke to who runs AI at a leadership level had hit this exact problem. Except almost none of them named it.

I think 2027 is when products in this space start to show up, and I think they will matter more than they currently look like they will.

That’s what stuck this month. If something here landed with you, or you think I’ve got it wrong, I’d love to hear it.