AI Agents help me manage outside boards, and a critical Claude limitation

I sit on three outside boards in addition to running Markup AI — Techstars, Alumni Ventures, and Citizens & Scholars. Each is a serious commitment: real fiduciary responsibility, hundreds of pages of pre-read per meeting, distinct organizational dynamics, and the expectation that I show up prepared to add value rather than just attend.

I used to prep for these meetings the old-fashioned way. Read the board book before the meeting, on the plane, etc. Take notes. Hope I remembered what happened last quarter. It was fine. It wasn’t great.

Recently, I set up an AI agent for each board. Here’s what I did, why, and how. Hopefully this becomes a standard practice that Brad, Mahendra, and I can write about in a future edition of Startup Boards.

The What

Three dedicated AI agents, one per board, each living in its own Claude Project. Each agent has access to that board’s Google Drive folder — where the board book, financials, minutes, charter, and other formal materials live — and that board’s folder or label in my mailbox, where the CEO, Chair, and fellow directors actually communicate between meetings. Each has a tailored instruction set defining my role on the board, the board’s strategic context, my fiduciary obligations, and a structured workflow for how I want it to serve me.

When I ask “what do I need to know for next week’s audit committee meeting?” the agent reads across both sources, synthesizes what’s on the table, drafts the sharp questions I should be asking, surfaces red flags in the materials, and tells me whether anything contradicts what management said last quarter. That’s the job.

The Why

The gap between an okay board member and a great one is usually preparation, continuity, and pattern recognition. The great ones come to meetings with context the rest of us are missing. They remember what was committed two quarters ago. They notice when this quarter’s narrative doesn’t match last quarter’s numbers. They ask the question that unsticks the whole conversation.

Doing that across three boards while running a company is genuinely hard. Materials accumulate. Continuity gets lost. By the time the next meeting rolls around, I’ve read a hundred other documents and the precise commitments from last cycle have blurred. The dynamic patterns — who pushed back on what, where management was evasive, what the Chair signaled — get fuzzy fast.

VCs are, in theory, the professional board members of the world — paid in part to sit on a dozen portfolio boards and do it well. I’m sure they’ve all figured this out already. The rest of us have to work at it.

This is exactly what AI agents are good at: pattern matching across a defined corpus, tireless recall, structured synthesis. They’re not replacing my judgment — that’s still mine, and should be. They’re giving me better inputs to apply judgment to, and freeing me from the parts of prep that are mechanical rather than directorial.

There’s a second-order benefit I didn’t expect. The exercise of writing down what “good prep” looks like — what I want the agent to produce for each meeting, in what order, with what lens — has articulated my own director playbook more clearly than I ever had before. Building the agent is itself a sharpening exercise. It also helps sharpen my leadership of my own board…but that’s a different story and true of almost all aspects of being an independent director on someone else’s board.

The How

A few design choices worth explaining.

Projects, not custom code. I considered building a custom application — I’ve been doing a lot of vibe-coding lately on other projects — but for this use case it would have been over-engineering. Claude Projects gives you exactly what’s needed: a stable knowledge base, custom instructions, conversation continuity. I had three projects set up in an afternoon. A custom app would have taken a week and ended up less capable.

One agent per board, not one combined. This was deliberate. Each board has different governance contexts, different confidentiality regimes, different people, different histories. Trying to combine them would have created confusion and real confidentiality risk for no benefit. I want my Techstars agent to know everything about Techstars and nothing about Citizens & Scholars.

Live knowledge sources, not uploads — with one important caveat. All the materials already live in folders I maintain in Google Drive and Email. Rather than uploading documents into each Project’s knowledge base, I have the agents query those folders live. The agents always see the latest version, I don’t have to remember to re-upload after a board book gets revised, and there’s a single source of truth.

The caveat is worth knowing before you set this up: as of mid-2026, Claude’s Google Drive connector reads native Google Doc, Sheet, and Slide files cleanly but sometimes stumbles on PDFs — which are the dominant format for board books, financial packs, and most legal documents. For boards where the substantive material arrives mostly as PDF, that’s a meaningful gap. Claude’s Google Drive connector can also only be set up to a single Google Drive, so if you’re like me and manage multiple drives (work vs personal), that can be a problem in data retrieval.

The workaround for both is siilar, which is to use a combination of the Google Drive connector and the Filesystem conncetor. Filesystem reads local files on your machine. So, if you set up Google Drive for Desktop to sync your board folders as a local mirror, then point the agent at the local copy through the Filesystem MCP server rather than the Drive connector, it’s reading local files. That recovers full access to PDFs and everything else. And you can simultaneously set up a Google Drive connector to a second drive. The cost: Filesystem only runs in Claude Desktop on the machine where the folders are synced. You lose the cross-device portability that a pure Drive-connector setup gives you. For board prep specifically — where I’m at my desk before meetings anyway — the trade is acceptable, also true if you really only have one device for the bulk of your work. If your prep routinely happens on multiple devices, weigh accordingly.

Tight email scoping. This was the trickiest piece. My personal email holds correspondence for all three boards plus everything else in my life. Each agent needs to see only its board’s email. The fix is two layers: I keep clean Outlook folders, one per board, manually and rule-based filed (same for Gmail labels); the Project instructions explicitly constrain email searches to the named folder plus a listed allowlist of expected senders. Without that scoping discipline, the agents would technically have access to my whole mailbox — which doesn’t work well.

Detailed instructions. Each Project’s instruction set runs to several thousand words. It covers who I am and how to calibrate for me (I don’t need governance fundamentals explained), the specific board’s context (mission, strategic priorities, key people, current decisions, known sensitivities), my fiduciary frame and cross-board conflict-watch obligations, a structured prep workflow — materials inventory, read priority, strategic synthesis, continuity check, sharp questions, red flags, my prior positions to revisit — committee-specific lenses, between-meeting workflows, output style (direct, opinionated, citation-required), and boundaries around confidentiality. Claude helped me draft this.

The detail is the point. A generic “help me prep for my board meeting” prompt gets you a generic answer. A few thousand words of context produces something useful.

Diagnostic before reliance. Before trusting any of these for real prep, I ran each agent through a diagnostic: a structured first message that exercises the Drive read, the email read, the scope discipline, and cross-source synthesis. If anything came back thin, generic, or missing citations, I knew the plumbing wasn’t right and I fixed it before relying on the agent’s output. I’d recommend this step to anyone setting up a similar system. The risk with AI agents isn’t that they say nothing — it’s that they say something plausible and wrong.

What I’ll Know Soon

I’ll know whether this actually changes my prep after a cycle of each board’s next meeting. Whether I’m asking sharper questions in board meetings, catching things I would have missed, doing better continuity across meetings. Whether the agents stay useful as materials accumulate, or whether I have to keep refining the instructions. Which parts of the prep workflow get used a lot and which get ignored.

For now, if you’re on outside boards and you’ve been wondering whether AI is mature enough to help with director prep, my early take is yes — but only if you’re willing to put real thought into the setup. The agent is only as good as the context you give it. Set up the knowledge sources cleanly, write the instructions like you mean it, and verify before you rely on it.

The technology is ready. The question is whether you’re willing to invest the hour or two to make it work for you.