Here's how I'm building my context library

I told you last week I'd make a decision about Obsidian. I'm not going with Obsidian — I'm going with Tolaria. It's an open-source macOS desktop application for managing Markdown-based knowledge bases, and it has a Claude Code integration baked in. The goal here is building a proper context library for Claude: organized Markdown files that give it real, persistent knowledge to work with.

There was just one problem.

The Mac app froze every time, right at the vault creation page. The GitHub build did the same thing. I nearly gave up, but then I looked closer. Inside the application folder, Tolaria stores the vault config as a JSON file — and that's when something clicked.

The freeze is in the vault creation flow, but the vault itself is just a file. So I skipped the flow entirely. I asked Claude Code to read the GitHub bug report, then write the vault JSON directly into the application folder. Ten minutes later, Tolaria launched clean, and everything else is really good.

If you're already deep enough into Claude workflows to care about a context library, you're probably comfortable enough with Claude Code to try this.

Last week I told you models are solved and context is the frontier. I want to add to that this week.

Building tools is solved too.

The cost of shipping software is down 99 percent. Apps that took months to build now take days of agentic engineering (I refuse to call it vibe coding anymore).

Human plus AI is just how it works now. You do the thinking, AI does the typing.

I built a private equity dashboard for a client recently. A few hours here, a few hours there. A few days total. The same dashboard would have taken months two years ago.

And it's not just the building, it's the fixing. Tolaria froze on me. Ten minutes later it was solved. When an open source app has rough edges these days, those rough edges are temporary — you can ship a workaround and move on.

There's a pattern emerging with the tools I'm finding myself drawn to, and Exo, an email client by Ankit V Gupta, fits it perfectly.

It's open source, you bring your own Claude API key, and then it works like Claude Code but for your inbox.

I cloned the repo. Played with it. Built a couple of small things on top. Then I stopped.

Why? It's still pre v1. The desktop experience is sharp. The mobile experience does not exist yet. I do most of my email on my phone.

So, I went back to Superhuman. The new voice features pull a lot of weight there. I can dictate "send him a calendar invite, then reply to this thread" and Superhuman handles it.

But I'm watching Exo. The day there's a clean mobile app, I'll plug my managed agents in as extensions. Imagine my SDR, AE, and account-manager agents working inside my mailbox, on my behalf, on every thread that needs them. That's the destination.

Email is the most-used channel in business. Whoever ships that future first will be huge.

My workflow has changed shape this year.

About 95 percent of my work happens inside Claude Code, running through Warp. I keep eight or nine tabs open at once, sometimes ten.

It's not multitasking. It's parallel work. One tab is drafting a proposal. Another is researching a deal. Another is summarizing a meeting. I check in, give a nudge, move on.

That's what made the next thing possible.

If you have the Claude Mac app, open it. Click the Code tab. You'll see a section called Routines.

This is how you turn a chat into a scheduled agent.

The flow:

Two routines I love right now:

Mercury makes this easy. It's a digital bank built for companies. Connect it once, the routine pulls the data directly. Not on Mercury? Plenty of workarounds. Schedule statements to your inbox and have the routine scrape them. Or upload weekly exports yourself. The data has to land in front of the agent somehow. The automation does the rest.

If you're not sure what to automate, try this:

Whatever Claude gives you back is the prompt for your routine. Plug it in, and let it run.

Every Friday afternoon, I spend twenty minutes scrolling through the week's Claude and ChatGPT chats. Some are throwaway. Some are gold. The gold ones get turned into something that runs without me.

Here’s how:

Pair this with the new Memory in Claude Managed Agents, which launched on April 23. Your agent now stores memories as files you can read, edit, and steer. The combination of Memory plus Routines plus your week's chat history is a flywheel.

Most of us repeat the same work all week without realizing it. The Friday review is how you spot it.

KiloClaw — A managed instance of OpenClaw. Their servers, their dashboard, their security and permission settings. Think of it as OpenClaw without the headache of running it yourself. I tried hosting my own OpenClaw before. The maintenance bled my time. Now KiloClaw runs it for me. I just got 60-day access. I'll report back.

getviktor.com — A Slack-based agent for teams already living in Slack. Worth a look if your company runs on it.

Anthropic Memory in Managed Agents — Mentioned above. Worth opening console.anthropic.com just to see it.

The five-year-old solution — I print my daily and weekly priorities every morning. The papers land in my office printer. My daughter is almost 5, and last week I realized she's exactly the right age for her first job: walk to the printer, grab the papers, put them on my desk. She'd take it very seriously. She'd probably want a title.

How can I automate more? That's the question. Can I involve my kids? Maybe. Nicely.

Three open webinars over the next two months. All at 5:30pm Turkey time (2:30pm UTC).

Registration links go live this week via LinkedIn and YouTube events.

Follow the AI Operator LinkedIn and YouTube pages and you'll get the links the moment they go up.

What do you want me to cover next? Hit reply.

Thanks for reading,

Tim