#010 - 35 hours

April 3rd 2026

I have more leverage than ever. And yet, I’m still the bottleneck for every prompt.

It’s been a few weeks since the last issue. I moved back to Canada, got settled, set up a Starlink so I can work from pretty much anywhere. Life stuff. Thanks for your patience! I have a backlog of things to write about, so expect a tighter publishing cadence over the next few issues.

Let's dig in.

A typical day these days: 8 or 9 terminal tabs open. Each one with a different Claude Code session. Each session, a different project.

Tab one: fixing a webhook signature verification for Tether. Tab two: closing SEO tickets for a client’s real estate site. Tab three: debugging a subscription billing edge case for PHAiTO. Tab four: building a product management dashboard in Laravel. Five through eight: more of the same. Different clients, different codebases, different problems.

Every tab needs me. Not to write the code; AI handles that now. But to provide the context. To explain what the project cares about, what the stakeholder cares about, what I care about. To review the output. To decide what’s next.

I shipped across six projects that week in about 35 hours. That sounds productive and it was. But it feels like being a switchboard operator.

AI gave me more leverage than I’ve ever had. Somehow, I’m still the bottleneck for every prompt.

The leverage is real

I want to be clear about something: this isn’t a complaint. The leverage is extraordinary.

A year ago, that same week would have been 80+ hours. Or more likely, I would have dropped two projects entirely because there wasn’t enough time. AI changed the math. I don’t write most of the code anymore. I describe what needs to happen, review what comes back, and steer when it drifts. The execution layer got compressed.

The Pragmatic Engineer surveyed 900+ engineers recently. 95% use AI tools weekly. 56% report doing 70% or more of their engineering work with AI. (Am I weird for being at 99%?). Claude Code went from launch to the most-used developer tool in eight months. This isn’t early adoption. This is the new baseline.

I’m living those numbers. The output is real. The speed is real. The ability to maintain six projects simultaneously as a solo developer; that was impossible two years ago.

But faster isn’t the same as free.

The bottleneck shifted

Here’s what nobody talks about in the “AI makes you 10x productive” narrative: the cognitive load didn’t decrease. It moved.

I used to spend 80% of my time writing code and 20% thinking about what to build. Now I spend 80% thinking and 20% reviewing. The typing got automated. The thinking got harder. Good thing I like thinking.

Context-switching across six projects means holding six different mental models. Six different codebases. Six different client priorities. Six different “what matters right now” decisions. AI can execute any of them. But AI can’t decide which one matters most at 2pm on a Thursday. That’s still me.

Thorsten Ball wrote something recently that stuck with me. He compared the current moment to a world where oil suddenly became infinite. Every structure built on the assumption that oil is scarce; pipelines, pricing, geopolitics; would need to be rebuilt from scratch. Code is going through the same thing. We built our entire industry on the assumption that code is expensive to produce. Now it isn’t. And most of us are still organized around the old assumption.

I’m organized around the old assumption. I’m in every terminal tab because that’s how a solo developer works. You do the thing. AI made me faster at doing the thing. But I’m still the one doing it, project by project, tab by tab.

That’s the trap. You get faster, so you take on more. You take on more, so you’re busier. You’re busier, so you feel productive. But you’re not freer. You just have more tabs open.

What has to change

Implementation cost is approaching zero. Everyone can feel it. So the popular take is that ideas become the bottleneck. I don’t think that’s quite right either.

I have no shortage of ideas. What I’m short on is systems. The ability to turn those ideas into running projects without being the one holding the context for every single one.

Right now I’m the driver of every bus. I need to be in the dispatch room, watching routes and only stepping in when something needs rerouting.

There are dozens of AI orchestration tools out there. Most of them are generic. Plug in your LLM, define some workflows, let agents run. They solve the execution problem. But execution was already mostly solved by tools like Claude Code.

The harder problem is the context problem. How does an agent know that this client cares about performance over features? How does it know that this codebase has a quirky test setup that breaks if you touch the config? How does it know that the PR should follow this project’s conventions, not generic best practices?

That’s the layer that’s missing. Not smarter models. Not better orchestration. A system that holds project context deeply enough that agents can make good decisions without my prompt every time.

The honest middle

I’m writing this from the messy middle. I have vision for where I want to take all of this. But reality hasn’t caught up yet.

I’ve been building a platform for my own practice. Something that holds project context, holds memories, gives me observability into agent work, coordinates agents, and eventually lets them pick up work without me prompting every step. Most of my available time in March went into the foundation: building the layers that agents need to operate autonomously.

I’m building a system that lets me go deeper on each project instead of wider. If agents can carry the routine execution; tickets, bug fixes, PR cycles; then I can spend more time on the work that actually compounds. Architecture decisions. Client relationships. Product direction. The stuff that needs a human with context, not a human with a keyboard.

And eventually, maybe, packaging that system so other builders can do the same.

Cool stuff from the internet

WP Fusion: We Built 4 Tools to Fix Our Support Pipeline

wpfusion.com

Jack Arturo documented how a two-person WordPress plugin company automated seven support processes with AI. Chrome extension that drafts support replies with full customer context. Auto-generated GitHub issues from tickets. Documentation screenshots that update themselves. The tools are free. The real story is a small team using AI to build exactly the kind of internal systems that matter right now.

Counselors by Aaron Francis

github.com/aarondfrancis/counselors

CLI tool that dispatches the same prompt to multiple AI coding agents in parallel and collects independent perspectives. Not task parallelism; opinion parallelism. Ask Claude, Codex, and Gemini the same architecture question, get three independent answers, synthesize. I’ve been using it for code review and it catches things a single-model pass misses.

Karpathy’s Guidelines for Coding Agents

github.com/forrestchang/andrej-karpathy-skills

Four principles distilled from Andrej Karpathy’s observations on where coding agents fail: think before coding, simplicity first, surgical changes, goal-driven execution. Packaged as an installable Claude Code skill. The “simplicity first” rule alone would fix half the AI-generated code I review.

What I’m shipping

A new task system. Built as a Laravel package with a Filament dashboard and CLI interface. Kanban board, inline editing, search, sub-project hierarchy. The point: agents and humans can both read and write to the same task list. I assign work, agents pick it up, progress is visible in one place.

Tether beta is live and has its first users. We’re working on expanding our carrier partnerships to offer better coverage. The product works. Now it’s about reach.

I’m also looking at open-sourcing some of the Laravel packages from my ops system. There are over 15+ internal packages powering the platform. Some of them; the connector libraries, the AI SDK wrappers; would be useful to other builders. More on that soon.

The AI productivity story most people tell is about speed. Ship faster. Code more. 10x output.

I’m less interested in that story now. The speed is table stakes. Everyone has it or will soon.

The story I care about is the one where a solo builder goes deeper on every project, not just wider. Where the systems handle the routine so you can focus on the work that actually matters. Not grinding harder. Not spreading thinner. Just building better.

I’m not there yet. But the direction is right.

Keep shipping,

Joey