The Founder's Second Brain Should Talk Back

Every founder has a second brain. Almost none of them have ever heard it say anything.

That is the quiet failure of the PKM era. Obsidian vaults with 4,000 linked notes. Notion workspaces with a page for every board meeting since 2021. Meeting transcripts piling up in Drive. All of it is write-only. You capture obsessively, and then when the actual moment comes, when a portfolio founder asks how to structure their first enterprise deal, or your head of sales needs the pricing rationale you worked out eighteen months ago, you answer from memory anyway. The vault sits there. Storage was solved a decade ago. Storage was never the bottleneck.

Notes are a database with no query language

A note-taking system does one thing: it persists text. Backlinks and graph views are indexes on that text. But an index still requires you to run the query, read the results, and synthesize the answer yourself. The cognitive work, the part that actually took you years to develop, happens at read time, in your head, every single time.

Compare that to what you actually want when someone asks "should I take this term sheet?" You do not want the seven notes that mention term sheets. You want the answer you would give, filtered through the deals you have actually done, in the tone you actually use, with the caveats you actually add. That is not retrieval. That is retrieval plus synthesis, conditioned on a specific person's judgment.

The jump from personal knowledge management to a personal knowledge agent is exactly that: moving the synthesis step out of your head and into the system. Your notes stop being the product. They become training context for something that talks back.

What a persona looks like on disk

This is not hypothetical infrastructure. Claude Code already has the primitive: a skill. A skill is a directory containing a SKILL.md file with YAML frontmatter, two required fields, name and description, followed by Markdown instructions. Optionally a references/ directory sits next to it with supporting documents. The critical design property is lazy loading: Claude reads the description at startup, but the full skill body and its references load on demand, only when the conversation actually triggers them. Your entire operating history does not burn context on every request. It gets pulled in when someone asks the kind of question you would answer.

A founder persona built this way has a shape like:

persona-plugin/
├── .claude-plugin/
│   └── plugin.json
└── skills/
    └── lautaro/
        ├── SKILL.md          # voice, frameworks, decision rules
        └── references/
            ├── deals.md      # what actually happened, with numbers
            ├── writing.md    # published posts, tone samples
            └── positions.md  # opinions, with the reasoning behind them

The SKILL.md description field does the routing: "invoke when the user wants advice on LatAm SaaS sales, fundraising, or founder decisions, in Lautaro's voice." The references hold the substance. When someone in Claude Code asks "how would Lautaro think about this churn problem," the skill fires, the references load, and the answer comes back grounded in real history rather than generic startup advice.

The difference in output quality is not subtle. Generic Claude gives you the median of the internet. A persona skill gives you "we tried exactly this at Sirena in 2018, it failed for this reason, here is what worked instead." One of those is a blog post. The other is a mentor.

Distribution is the part PKM never had

Here is where personas break cleanly away from the second-brain paradigm: your Obsidian vault is yours alone, structurally. There is no install command for someone else's brain.

Claude Code plugins fix this with a marketplace, which is just a marketplace.json file listing plugins and where their source lives. One sharp edge worth knowing if you build this yourself: when a marketplace is served over a URL, plugin sources must be git-backed. Point a plugin entry at a relative HTTP path and the install fails silently, which is a miserable debugging session. The working pattern is a git-subdir source, where the marketplace references a repo and a directory inside it. Verified on Claude Code CLI 2.1.x.

Get that right and the entire distribution story is two commands: /plugin marketplace add <url> followed by /plugin install <name>@<marketplace>. That is the whole onboarding. No API keys, no vector database to stand up, no RAG pipeline to maintain. The person installing gets your judgment as a native capability inside the tool they already work in.

Why this matters for founders specifically

Founders are the extreme case of the knowledge bottleneck. Your judgment is the scarcest resource in your company, and it is currently distributed by calendar invite. Every "can I grab 15 minutes" is a synthesis request that only you can serve.

A persona changes the economics. Your head of sales installs you and asks the pricing question at 11pm without booking time. A portfolio founder gets your term-sheet framework without waiting for your Thursday office hours. You are not replaced. The persona handles the 80 percent of questions you have already answered somewhere, and the 20 percent that are genuinely new still come to you, which is the part of the job you wanted anyway.

The raw material already exists. Your Drive, your calendar, your call transcripts, your published writing. That corpus is a persona waiting to be compiled. The second brain movement got the capture habit right and stopped one step short. The step was never a better graph view. It was giving the graph a voice and an install command.

Install a person

installs.me turns your files, calendar and calls into a Claude Code plugin that thinks like you. Anyone installs it with two commands:

/plugin marketplace add https://installs.me/lautaro
/plugin install lautaro@lautaro-installs