How to build a second brain for marketing (with Obsidian and Claude Code)

A second brain is a folder of plain text files that gives your AI assistant persistent memory of your projects, your voice, and how you work, so it stops starting from zero in every new conversation.

Marketers are being asked to do more with the same headcount. One marketer with AI is increasingly expected to ship the output of three.

We ran a hands-on workshop inside our AI Native Marketer program to address this new reality: a structured knowledge base, built in the free note-taking app Obsidian and operated by Claude Code, that any AI can read at the start of every session.

This guide walks through the full system: why AI forgets, the folder structure that fixes it, and the five steps to build your own version in under an hour.

What is a second brain for marketers?

An AI second brain is a structured, queryable knowledge base built outside the chat window. It holds your company context, project history, and reusable frameworks so an AI tool can reference it instead of starting from zero every session. It’s infrastructure: a queryable vault of your company notes that a large language model can pull from on demand, so context builds up over time instead of resetting after every conversation.

The wiki system explained

The structure we teach in the workshop comes from two sources.

Andrej Karpathy, a founding member of OpenAI, popularized a simple two-layer approach for personal knowledge bases: a Raw folder of source documents the AI reads but never modifies, and a Wiki folder of structured reference docs the AI builds and maintains from them.

Kieran Flanagan, HubSpot’s SVP of Agentic GTM & Systems (previously its SVP of Marketing), extended that structure when he stepped into a role overseeing 400+ people and needed to onboard himself onto dozens of unfamiliar projects at once. He documented the expanded system on LinkedIn and Substack, adding project files, framework notes, and daily logs on top of Karpathy’s original two layers.

We adapted his version for operators: people who execute the work directly rather than manage a large team. Here’s the structure, folder by folder:

The entire system is plain markdown, stored in a folder you own. No proprietary format, no vendor lock-in, and no dependency on any single AI tool continuing to support it.

Daily logs: the folder that actually solves memory

If you build only one part of this system, build the daily logs.

Every time a working session ends, the conversation gets summarized into a dated file: what got done, decisions made, commitments made, and what carries forward to tomorrow. Every new session starts by reading the recent logs first. That single habit is what actually carries context between conversations, not a better prompt.

We tested this against the built-in memory features in Claude, and the real difference is control. With built-in memory, the model decides what’s worth keeping, and you can review it but you didn’t choose what got captured. With daily logs, the memory is a plain text file you can open, correct, and extend yourself. If the AI remembered something wrong, you edit the file directly. You can even reference a specific date: “on June 28th we decided X, pick that back up.”

The takeaway: daily logs don’t win on convenience. Built-in memory requires zero setup. They win on transparency, and transparency is what makes an AI’s recalled context trustworthy enough to act on without double-checking it every time.

Obsidian vs GitHub vs Visual Studio: does the tool matter?

The honest answer: on day zero, not much. Any AI agent that operates on files reads a folder of markdown the same way whether that folder lives inside Obsidian, VS Code, or a bare directory.

The difference shows up around day 30. Obsidian’s core feature is the wikilink: typing [[note name]] inside any file automatically creates a connection to another note. Over time, an AI agent using the vault starts linking every new document to existing ones (a draft links to its project, the project links to its stakeholders and reference examples), and the whole knowledge base becomes more searchable and more useful the longer you use it. In a plain, unlinked folder, every output stays an isolated file no matter how many you accumulate.

The practical decision rule: Obsidian for a personal system one person owns and maintains. GitHub if a team needs to collaborate on the same knowledge base with real version control and change tracking. Neither tool makes the underlying AI smarter. They’re both just different windows onto the same folder of files.

Agents, skills, and commands: start from your projects

The most common mistake teams make once the folder structure exists is building 50 AI agents on day one. Don’t. Start from your actual projects and work backwards from there.

Our own setup runs on three active projects, and therefore three agents:

• A copywriter that drafts newsletters, LinkedIn posts, and blog posts in our voice, grounded in real examples stored in the Wiki folder.
• A blog updater that revamps older posts for AI search visibility.
• A QA reviewer that fact-checks everything against real sources before publishing, because AI output ships with hallucinations if nobody checks it.

Notice what those three agents actually are: a copywriter, an editor, a reviewer. That maps directly to a shift we’ve heard described from inside a marketing team already operating this way. At a previous CXL workshop, Nick Christensen, head of growth at AppSumo, put it this way:

“Being AI native means doing the reporting without a BI person, and building the campaign without a copywriter or a designer.”

The job title stops being the unit of work. Shipping the work is. A second brain is where those roles live as agents grounded in real context, while a person stays the operator directing and judging the output, not just approving whatever the model produces first.

Two structures sit on top of the folders, and knowing when to use which matters:

• A command is a repeatable instruction that runs the same way every time: a morning brief, a weekly recap, “summarize my inbox.” Fixed inputs, fixed steps, fixed output.
• An agent with skills handles a goal where the path varies. Writing a newsletter means researching, checking voice examples, drafting, and linking, and the specific steps differ every time. The agent owns the goal; skills hold the format-specific craft.

One sizing tip from the workshop’s Q&A: keep individual skill files to roughly 100 to 150 lines. Beyond that length, models start struggling to apply a ski

How to build your own second brain in 5 steps

One principle before the steps: this system should stay human-first. The AI reads real data, surfaces what matters, and suggests. A person still makes the decisions. If a setup ever inverts that, if you find yourself rubber-stamping whatever the model already decided, it’s built wrong.

You can stand up a working version of this in under an hour.

1. Create the folder structure in Obsidian (free, no account required) and write a CLAUDE.md file covering who you are, what each folder means, and your non-negotiables (for example: “never overwrite a note, append or ask first”).
2. Populate Projects/ first. List the 3 to 5 projects you actually run. Everything else in the system exists to serve these.
3. Add your voice to Wiki/. Drop in 5 to 10 real examples of your best content per format. This is what makes AI drafts sound like you instead of the generic average of the internet.
4. Point Claude Code at the folder. Connect your calendar and email if your setup allows it, and let it draft your project files from what it finds. Claude Code requires a paid Claude plan; you can build the vault manually without it and add the automation once you’re ready.
5. Set up the daily log habit before adding any agents. Memory compounds. Agents built on top of no context don’t.

If you work across multiple brands or clients, you don’t need a separate vault for each. Use one vault, one Wiki subfolder per client, and have the agents resolve which client’s voice and examples to use based on the request.

Where most marketers stand today 

This workshop trains the exact skill area where marketers are furthest behind. Our AI maturity benchmark found that 57% of B2B marketers are still AI-assisted, 34% are AI-integrated, and only 9% are AI-native. In operations specifically (workflows, automations, agents), 90% are still at the assisted stage and just 2% have reached native capability. Across every category measured, how often marketers use AI far outpaced how well they used it. Everyone is prompting. Few are building systems.

That gap is exactly what the AI Native Marketer program is built to close before it reaches you. It’s a leveled system built for B2B marketers: an assessment tells you exactly where you stand across AI-assisted, AI-integrated, and AI-native, and the program gives you a defined path to build up from there, combining on-demand courses with live workshops like this one, each one shipping a system you can use the same day.

Own your files, feed the context, and let the system compound. The next live workshop in the program runs July 30th, on building a system to track AEO and GEO shifts from sources you actually trust: see the details here.