If you run or work in an agency, manage clients, or own anything more than a single workflow, you know how the first 45 minutes of the morning vanish into Gmail, Slack, Fireflies, the CRM, and whatever doc was last on screen, just trying to remember what mattered the day before. 

It used to be I’d be thinking about a pricing decision I owed the agency team, a roadmap call for my habit-tracking app, three Slack threads I’d half-read on my phone the night before, and a sales follow-up that needed to go out today.

Those days are over. About six months ago, I rebuilt how I work using Claude Code as the engine of what I’ll call my second brain. The Monday morning synthesis now takes about a minute.

Here’s what I built, why, and how you can do the same.

Why most second-brain setups break down

The “second brain” idea isn’t new. Tiago Forte’s “Building a Second Brain,” the PARA method, Notion, and Obsidian were all built on the same instinct: externalize what you’d otherwise try to remember.

Capturing information works. Recall mostly works. But the bit that adds value is what happens after recall: turning what’s stored into something you can act on.

Three failure modes show up in almost every implementation I’ve seen:

• Passive storage. Information goes in. It sits there. The only way out is search plus your own memory of what you tagged it. Meeting notes are a perfect example.
• Context-switching tax. Even when you find the right note, you’re still copy-pasting and re-prompting your way to a usable output.
• No action layer. A second brain that can’t draft, retrieve, or execute tasks can end up giving you more work than you started with. Over time, it builds up into an excess of notes that just leads to cognitive overload.

Documenting tasks isn’t the problem. The problem is that the documents and conversations are scattered across a hundred different apps, and nothing reads across them.

What’s missing is the layer on top: something that can pull from all of it and actually do the work.

Dig deeper: How to turn Claude Code into your SEO command center

How Claude Code changes the equation

Most general-purpose AI assistants are stuck behind a chat window. They can answer questions, but they can’t reach into your file system, remember what you said last week, or do anything outside their own UI.

Claude Code (Anthropic’s developer-focused agent) ticks four boxes that, together, change how you can operate day to day:

• Native file system access: It reads and writes inside an actual project folder, accessing your local files like any directory.
• Persistent, structured memory: It remembers across sessions through Markdown files I curate.
• MCP integrations into the tools you already use: The Model Context Protocol lets it plug directly into Gmail, Slack, Google Drive, HubSpot, and Scoro. No data migration. No rebuilding workflows around it.
• An action layer: It drafts internal docs, prepares analysis, and handles the repeatable parts of my workflow.

The last one is the most useful: moving from something that just stores information to something that acts on it. This is where I can save most of my time (and you can, too).

The four layers of an AI second brain

The architecture of my second brain is organized into four layers, stacked on top of each other.

1. Memory 

This is contained in a small set of plain Markdown files. One describes me and what I work on. One holds running facts worth keeping, such as pricing decisions, client preferences, and retainer structures. One defines the personality I want it to take when working with me.

These load automatically every session, so I never re-explain context.

The interesting bit is that the memory grows by itself. Every conversation gets dumped into a daily log. Once a day, a small job runs over that log and decides what’s worth promoting into long-term memory.

Throwaway chat doesn’t make the cut, and over six months, this has built a surprisingly accurate model of how I work and what each client needs.

2. Search

The long-term memory file stays small on purpose. So every daily log gets indexed into a local database.

When I ask something specific like, ‘What did we agree with this client on internal linking back in February?’ it pulls the actual conversation and context to review.

3. Skills

A skill is a focused capability I’ve defined once and can invoke by name: draft a client brief, build a sales proposal, reply to an email in my voice, summarize a discovery call into a scope of work.

Each skill is small, single-purpose, and inherits the memory layer underneath.

For example, the ‘draft email’ skill isn’t generic because it knows my tone, the recipient, and what we last discussed, by training on all the data it reviews day to day.

I deliberately avoided one giant do-everything agent because what I really wanted was an advanced assistant that can work on small, composable skills that improve each day to help support me while I work on the bigger decisions.

4. A heartbeat

Every hour during the working day, a small process pulls a snapshot across my tools, such as new emails, calendar shifts, Slack threads I’m in, and HubSpot pipeline movement, and decides whether any of it needs me.

If it does, I get a Slack ping with a one-line summary and, where appropriate, a starting draft for me to work from.

Dig deeper: How a ‘client brain’ gives AI the context SEO work needs

Where it pays back hours every week

Here’s where the time actually comes back:

Faster context-gathering for client work

When a client emails for a status update, the second brain has already pulled together every Fireflies transcript, Slack thread, and HubSpot deal note for that account.

What used to be 30 minutes of searching before I write back is now 30 seconds of context, so I can spend the time on the reply itself, not the digging.

Faster data analysis

Whether it’s analytics, rank-tracker data, or Search Console information, the second brain can pull what I need with context for review.

Discovery to scope

New retainers used to take days of back-and-forth. Now the second brain reads the discovery transcript, the email thread, and historic data on similar engagements, and it produces a scope I can pressure-test rather than build from scratch.

All of it adds up to working more efficiently and giving clients better service because the important stuff isn’t slipping through.

---

The guardrails that make this work

Setting up a tool this powerful needs guardrails. Otherwise, you end up with an agent doing things you didn’t ask for.

Read-only by default

Every integration starts read-only. So for Slack, calendar, and Gmail, it can see and draft, but nothing goes out without me.

Write access gets added one tool at a time (and only once I trust the outputs from that tool). An agent with broad write access is one prompt injection away from doing something you’ll regret, and that risk compounds the more tools you wire in.

Memory hygiene matters

The temptation is to dump everything into the tool. Don’t.

Long-term memory should hold things that change how the agent acts: client pricing, ongoing decisions, and ways of working.

Trust the draft, verify the action

Read every email it drafts before it sends. Read every brief before it lands with a client.

The point isn’t to remove yourself from the work. It’s to give you a head start based on what it knows, while you still use your own expertise to finesse the final decisions.

Dig deeper: How to train Claude to sound like your brand

How to build your own second brain

You don’t need my exact stack. Use the tools you’re most comfortable with, along with this process.

• Pick the four or five places real decisions live for you. Email, calendar, one messaging tool, your CRM, and your task tool. That’s enough.
• Add a transcript layer. Calls are where the most context gets lost in agency work.
• Build memory first. A ‘this is me’ file plus a daily log that gets distilled. Don’t move on until talking to it feels like talking to someone who knows your business.
• Add skills one at a time. Pick the most repetitive thing you do and build a skill for it.
• Add the heartbeat last. Once retrieval and skills both work, give it a schedule. Start with notifications only. Add write access slowly.

This is a second brain, but don’t let it replace your actual one

The goal of a second brain isn’t to replace your existing one. It’s to make sure you can be as efficient as possible with your day-to-day work, providing more value for your team and clients.

The tools to build this didn’t exist 18 months ago. They do now, and the time it takes to set up is paid back within a fortnight.

Dig deeper: How to build custom SEO reports with Claude Code and Google Search Console