#008 - From Doing to Directing

February 10 2026

I want to tell you about a shift that’s happening in how I work. Not a finished system. Not a “here’s my perfect setup” post. More like: here’s what actually works, here’s what doesn’t yet, and here’s why I think the direction matters.

Every morning, my AI assistant sends me a brief on Telegram. Calendar, priorities, what’s stuck, what shipped yesterday. I scan it over coffee and reply with a plan. By the time I sit down to work, the day has direction.

Every evening, it runs an end-of-day wrap-up. Logs what shipped, what’s still open, what moved to tomorrow. Confirms with me on Telegram before committing anything. My Obsidian vault stays current without me manually updating it.

Between those two bookends? That’s where it gets murkier, for now at least.

What Actually Works

Over the past few months, I’ve built over 70 small, focused skills for my AI assistant. Each one started as a conversation. “I need a way to process my reading highlights into my vault.” So we talked through it, tested it, refined it, and saved it. “I want code review that actually knows our codebase conventions.” Same process.

The skills compound. The daily brief reads from the task system, which syncs with my project tracker, which feeds into how the code review skill prioritizes what to flag. Everything connects.

Content drafting in my voice. Task management across projects. Meeting notes that auto-organize into my vault. Clipping processing. Tweet drafts. Monthly summaries for clients. These all work. I use them daily.

And because it runs in my terminal and on my server, not in a browser tab, it can do things a chat window can’t. Run tests. Commit code. Monitor errors. Send me alerts.

The morning brief and the evening wrap-up are the most tangible wins. My day starts focused and ends documented. That alone changed how I work.

The thing I keep learning is that the work that matters most isn’t building features — it’s building the system that builds features. Every solved problem should teach the system something. Every workflow should get a little better each time. The team at Every calls this “compound engineering”: each unit of work makes the next one easier. I’ve been doing a version of this instinctively, but naming it and formalizing it changed how I think about where to spend my time.

What Doesn’t Work Yet

I still manually kick off most tasks. “Review this PR.” “Draft that email.” “Run the SEO audit.” The skills exist and they work well, but I’m the one deciding when to run them. There’s no scheduler spinning up background agents while I’m focused on something else. Not yet.

The vision is a system where agents handle routine work automatically. Scheduled security reviews. Proactive error triage. Content pipelines that draft without me asking. I can see the architecture. I’ve built some of the pieces. But the orchestration layer that ties it all together? Still assembling that.

If you’re on X, you’ve seen the claims. Fully autonomous AI agent fleets. Ship while you sleep. I wanted that to be real. I tried it with OpenClaw — gave it real tasks, real codebases. The results weren’t there yet. Not for professional software development. What it produced was too sloppy. Too many assumptions, too eager to ship. Not enough awareness of the conventions and guardrails that make code actually maintainable. I even tried giving it my Claude skills, my vault, and other systems, and it still isn’t producing anywhere near the same results as me kicking things off in Claude Code. I’m still thinking that OpenClaw and systems like it are the future though. We’re just early.

I think the path forward isn’t “turn it loose and hope.” It’s building up the system with more guidance, more context, more guardrails. Teaching it how you work before asking it to work on its own. That takes time. And that’s okay.

This is the reality of building in public. You see the direction before the system is complete. And the temptation is to write about it like it’s already done.

It’s not. But the direction is clear.

Two Things That Just Landed

Two new Claude Code features showed up this week that push the “directing” side forward.

Auto-memory. Claude now automatically saves useful context as it works. Project patterns, debugging insights, architecture decisions, your preferences. It writes its own notes into a persistent memory directory. Next session, it picks up where it left off. No more re-explaining your project structure every conversation.

I’ve been building something similar manually for months. CLAUDE.md files, memory layers, handoff documents between sessions, Obsidian integration. Auto-memory does a version of this out of the box. It’s not as tailored as a custom setup, but the fact that it exists as a default feature tells you where things are heading. Memory is becoming table stakes.

Agent teams. This one I used for the first time today. Multiple Claude Code instances working together. A lead agent coordinates, spawns teammates, assigns tasks. Teammates work independently, each in their own context, and communicate with each other directly. Shared task lists. Real coordination.

I used it on a client project to implement a series of improvements to a sophisticated automation pipeline. Instead of one agent working through everything sequentially, the lead broke the work into parallel tasks and teammates executed simultaneously. Different files, different concerns, same codebase.

It’s experimental. There are rough edges. But the model is right: a lead that directs, workers that execute, a shared system that keeps them coordinated.

The Bigger Picture

I’m far from the only one seeing this.

Boris Cherny, the creator of Claude Code, shared his setup recently. Five Claudes running in his terminal. Five to ten more in the browser. 259 PRs merged in 30 days. His top tip: spin up 3-5 git worktrees, each with its own Claude session. That’s a developer who stopped coding and started directing.

Shubham Saboo called this the shift from “vibe coding to fleet command.” Modern AI developers manage agent teams, not prompts. Serialize by default, parallelize explicitly. Most people over-parallelize and get interleaved garbage. The discipline is in knowing when to fan out and when to stay focused.

The mental model is changing. Not “me and my AI assistant” but “me and my agents.” Plural. Each one focused. Each one disposable. The human stays in the loop at every decision point. The agents handle execution between those decisions.

Cool Stuff

NativePHP for Mobile is Now Free and MIT Licensed

laravel-news.com

If you’re in the Laravel world, this is big. Build native iOS and Android apps with Laravel. The core framework and nine plugins are now completely free and open source. I haven’t shipped anything with it yet, but I’m watching this closely.

Follow Your Passion (But Not Like That)

Arvid Kahl / The Bootstrapped Founder

Arvid nails something that took me years to learn. Don’t monetize the joy. Find people who share your passion and solve their problems. “Who else has this passion, and what challenges do they face?” That question is worth more than any business plan.

What I’m Shipping

Tether is in beta. Onboarding the first users. If you deal with SMS — travellers, solo founders, digital nomads, expats, remote teams — it’s built for you. tethermobile.com

Still balancing client projects alongside personal products. The agent setup makes that possible. Not easy. Possible.

The gap between “I want this to exist” and “it exists” keeps shrinking. Not because AI is magic. Because the plumbing around it is getting better. Memory that persists. Agents that coordinate. Skills that compound.

You don’t need to run an agent fleet tomorrow. Start with one thing that saves you ten minutes a day. Then another. Then connect them. The tools that compound are the ones worth investing in.

Keep shipping, Joey

P.S. What’s one thing you’ve automated with AI that surprised you? Hit reply. I’m collecting stories from other builders.

P.P.S. If someone you know would dig this, forward it along. They can subscribe at jkudish.com/newsletter.