Ask your meeting history anything: the CEO's query layer

Open Claude. Point it at your meeting corpus. Ask any question in plain English. Get answers in thirty seconds, with citations.

What you'll have when you're done

A query interface that sits on top of every meeting you've recorded. You open a terminal in the folder where your Granola transcripts live, type one command, and ask questions in plain English. What objections came up most in the last thirty sales calls? What commitments did I make in 1:1s last week that I haven't followed up on? Summarize every concern our chair has raised this year. Each question returns a structured answer with citations to the specific meetings it came from. Same interface for every query type: pipeline, hiring, board, customer themes, your own behavior in the room.

No new install. No SaaS. No script to maintain. The interface is claude code in a terminal. The corpus is the folder you already built in Granola → markdown. The queries are whatever you can phrase in a sentence.

Why this matters

Every CEO already does this work. Badly. Before the Wednesday board call, you scroll your notes app for the last board email. Before the Friday 1:1, you try to remember what your VP of Sales said about hiring three weeks ago. Before the customer follow-up, you replay the call in your head and write down what you think they said.

That's the same query the corpus answers in thirty seconds, with the verbatim source.

The query layer is the single highest-frequency tool in my stack. I run it ten or fifteen times a day. None of those queries existed a year ago, because the corpus didn't exist a year ago. Now the bottleneck on running my week is no longer my memory.

What you need first

• The Granola → markdown pipeline already running. This workflow runs on the markdown corpus the pipeline produces. If you haven't set it up yet, do Granola → markdown first. Thirty minutes. Comes back when done.
• Claude Code installed. If you've never installed it, start at claude.ai/code. Free tier is sufficient for this workflow.
• At least ten meetings in the corpus. The query layer is real after one meeting. The value only shows up at volume. If you set up the pipeline yesterday, run it for a week before judging the output.

Step-by-step

Step 1Open claude code in your notes folder

The single most important configuration choice is where you open claude code. Claude reads the files in the directory you launched it from. Launch it in your home folder and it sees everything. Launch it in your notes folder and it sees only the corpus.

Open your terminal and run:

cd ~/notes
claude

~/notes is the folder you set up in the granola → markdown workflow. If your folder is elsewhere, use that path. Once the prompt is ready, claude code is waiting for your first query.

Step 2Run your first query

Start with something specific that you know is in the corpus. Don't open with a sweeping question on day one. The point of the first query is to prove the citations work, not to surface insight.

Paste this and replace the bracketed bits with a real customer or topic from your transcripts:

Look across every meeting in ~/notes/granola/. Find every meeting where [customer name or topic] came up. For each, give me one verbatim quote and the source filename. Brief, no summary.

You should get back four or five quotes, each followed by the exact filename Claude pulled it from. Open one of the cited files. The quote should match the file. If it doesn't, that's the most important thing to know on day one (more in When it breaks below).

Step 3Move to the query gallery

Once the first query works, the rest of the workflow is muscle memory. Here are the queries I run weekly. Each one is paste-ready. Adjust the bracketed parts to fit your business.

Pipeline.

[VERIFY: Andrew to confirm or replace with his actual prompt]

Across every sales call in the last 30 days in ~/notes/granola/, what are the top five objections the buyer raised? For each:
- The objection in one sentence
- Three verbatim quotes (one per distinct call)
- The deal stage where it came up
- The source meeting filenames

Team and 1:1s.

[VERIFY: Andrew to confirm or replace]

Across every 1:1 I had this week in ~/notes/granola/, list every commitment I made and every commitment my report made to me. Format:
- Person · their commitment · my commitment · the deadline if mentioned · source filename
Skip anything that wasn't an explicit "I'll do X by Y."

Board and investor.

[VERIFY: Andrew to confirm or replace]

Read every meeting in ~/notes/granola/ from the last 90 days where our chair was an attendee. List every concern she raised, what I said in response, and whether the concern was resolved (check later meetings for follow-up). Cite the source meeting for each.

Hiring.

[VERIFY: Andrew to confirm or replace]

I'm comparing two candidates for the VP role. Find every interview transcript in ~/notes/granola/ tagged or filenamed with [Candidate A] and [Candidate B]. For each candidate, summarize how they answered:
- How they handle disagreement with a manager
- A failure they led their team through
- What they look for in a hire

Quote verbatim where the answers are most distinctive. Cite the source meeting per quote.

Strategy.

[VERIFY: Andrew to confirm or replace]

Across our five largest current deals in the pipeline ([deal A], [deal B], [deal C], [deal D], [deal E]), what's the recurring theme in what the buyer says they want? Where the theme is sharper in one or two of the calls, quote those verbatim. Source meeting per quote.

Personal.

[VERIFY: Andrew to confirm or replace]

Across my 1:1s this month in ~/notes/granola/, where did I dominate airtime? For each 1:1:
- Estimate the talk-time ratio (mine vs theirs)
- Flag any conversation where the ratio is worse than 60/40 in my direction
- Quote one moment in that conversation where I could have made space for them and didn't

There are more. Every query has the same three parts: a verb (find, compare, summarize, list), a scope (which meetings, which timeframe, which people), and a format (what to return). Once you've run six or seven, you'll feel the shape and stop looking up the templates.

Step 4Read the citations, not the answer

The most common mistake new users make is reading Claude's answer and stopping there. The citation is the proof. Especially in the first month: open one or two of the cited files for every query, find the quoted sentence in the source meeting, and verify it. If the quote doesn't appear verbatim in the source, the answer is unreliable, even if it sounds right. Claude can confabulate from a corpus. Citations are how you catch it.

After two weeks of spot-checking, you'll develop a feel for which kinds of queries Claude handles cleanly (factual recall, verbatim quote extraction, simple aggregation) and which are riskier (multi-step inference, vague summarization, behavior pattern claims). Lean on the former. Verify the latter.

Step 5Save your top five prompts

The query gallery above is a starting point. The actual gallery is the set of queries you find yourself rerunning. After the first week, copy your five most-used prompts into a single file called prompts.md in your notes folder. Then in claude code, when you want one of them, you can say "use the Friday commitment review prompt from prompts.md" and Claude reads the file.

That's the entire workflow. One terminal command. One prompt at a time. One corpus.

How you'll know it's working

The signal is the first uncomfortable answer.

Two or three weeks into running the query layer, you'll ask something you thought you already knew the answer to. What did our lead investor flag in the January call? The answer that comes back will be different from what you remembered. The citation will prove the answer is right and your memory was wrong.

That moment is the workflow.

A secondary signal: the day you stop scrolling your notes app to find a meeting. You used to scroll. Now you ask. When the scroll has dropped out of your week, the workflow is installed.

When it breaks

• The corpus is too big for the context window. Claude has a finite context. If you point it at three years of transcripts and ask one query, it may sample, miss things, or refuse. Fix: scope the query. Add a date range (in the last 30 days), an attendee filter (where Priya was present), or a meeting-type filter (only sales calls). Specific scope, specific answer.

• The query is too vague. What's interesting in my meetings? returns junk. What did the buyer at TechCorp say about pricing on the April 15 call? returns gold. The query layer rewards specificity. Treat every prompt like a question you'd ask a research analyst: subject, scope, format.

• The citations don't match the source. Claude says person X said Y, but X never said Y in the cited meeting. Fix: ask Claude to quote verbatim and reproduce the surrounding three sentences. If it can't, Claude hallucinated. Reframe the query and run it again. Persistent mismatches usually mean the corpus is too large for the context window (see first failure mode); reduce scope.

• Claude refuses or stalls on a complex query. Long, multi-part prompts ("find every X, compare to Y, then summarize Z") sometimes fail mid-answer. Break the query into two or three smaller queries and chain them yourself. The cost of three small queries is the same as one big one.

• The corpus is empty or stale. You set up the pipeline last week, ran it once, and didn't add new meetings. Run qmd embed from your notes folder. If you haven't logged meetings recently, the query layer can't help. Volume is the prerequisite. Get to thirty meetings before judging the output.

Where this fits in your harness

This is the query layer of the harness. Every other workflow in this batch is a specific, repeated, or scheduled version of a query you could run by hand here. It runs on top of the corpus the Granola → markdown pipeline produces. The full case for the pipeline and the twelve workflows lives in Granola for CEOs: the highest-ROI AI install of 2026.

The siblings most relevant to this workflow:

• Pre-meeting brief · the same query, scheduled to fire fifteen minutes before every meeting on your calendar.
• The commitment ledger · the "team and 1:1s" query above, turned into a weekly file you refresh.

Build the query muscle here first. The rest of the index becomes paste-ready prompts against the same corpus.