All-hands meeting agenda template

The block that makes this template different is the Q&A split three ways: asked, answered, and owed. Most all-hands records keep the slides and lose the questions, which is precisely backwards. The slides already exist as a file.

The owed table is the whole point and it has to have names and dates in it. An all-hands that generates unanswered questions and no list of them teaches the room not to ask, and the second all-hands after that one is silent.

What it decides. Nothing, formally. It commits. What gets said from the front is what the company has now said out loud about itself, and people plan their year around it whether or not it was meant that firmly.

Who is in it. Everyone, which means the record is for the sizeable fraction who could not attend: other timezones, people on shift, anyone who joined last week. It is the one meeting in this pillar whose audience for the record is larger than the audience for the meeting. What executives keep covers the narrative side of it, and what happens to the number that gets quoted.

What must survive it. The question that was asked and not answered, which is the fact that determines whether anyone asks next time. The number as it was actually stated, because it will be requoted for a quarter with increasing confidence and decreasing accuracy. And the commitment made from the stage, which nobody thinks of as a commitment at the time because it was phrased as reassurance.

What an all-hands record needs that others do not. The three-way Q&A split, so an owed answer has somewhere to live between meetings. Attribution of questions where the asker was public, and deliberate non-attribution where they were not. And a plainly stated one-message line at the top, because a record of an all-hands that opens with the numbers reproduces the meeting's own habit of burying the thing people came to hear.

Somebody posts the deck afterwards, and occasionally a recording, and that is treated as the record. Neither is one. A deck is the part that was already written down, and a full-length video is not something a person joining in six weeks will ever watch.

The Q&A is where this bites hardest. Questions come in from three places at once, the chat, the room and a submission form, and the last few get an honest "let me come back to you on that". Nobody is assigned to collect them, because collecting them is a job that belongs to whoever is currently on stage. So the promise is made in good faith and then simply evaporates, and the next all-hands gets fewer questions for reasons everyone can feel and nobody names.

This half is a product feature. Exactly what it does, in four steps:

1. The day is already recorded. Earkeep captures your microphone and your system audio continuously and transcribes on your device, and the audio is never written to disk at all. There is no start button and no bot joining the call. Always-on capture covers why, and private by design covers where the audio goes, which is nowhere.
2. You mark the meeting once it is over. Select the range on the timeline and save it as a meeting. An all-hands is always on a calendar and is usually the only event of its kind that month, so the connected event names and bounds the span cleanly. Calendar connection has the detail.
3. You pick a saved template in the composer. Settings → Templates holds instruction presets, each with a name, an agent, a working directory and the prompt text. Opening the agent panel on a saved meeting offers them, and the meeting and its transcript are attached to the prompt for you, as resources, so the preset itself is just the instruction.
4. The agent writes the file. It runs on your machine, as your own CLI, with your own key, in the directory the preset names. Agents and MCP explains how that works.

The "one click" is picking the preset. Nothing fires a template because a meeting's title matched a pattern; Earkeep has no such matching.

The preset for this page, with ~/notes/all-hands as the directory:

The distinction between an answer and a non-answer is the instruction worth being firm about. "We are looking at that" reads as an answer in the room and is an owed item on paper, and a record that cannot tell the two apart produces an owed list of length zero every time.

Earkeep has no scheduler. There is no nightly job to switch on, no "run this agent every week" setting, and there is not one hidden in a menu either. What it has instead is a directory of plain markdown files and a read-only MCP server on loopback, which is everything a scheduler you already own needs.

The pattern on top of that is a compiled wiki: raw sources stay immutable, an agent writes one page per concept from them, and a schema file states the rules. The LLM Wiki is the pattern and from what was said to what you know is the pipeline a meeting travels along it. For an all-hands the destination is a page per recurring theme plus one ledger of owed answers, because the pattern that matters across a year is which questions keep coming back, which is a fact no single meeting contains.

Keep the ingest instructions in a file so a scheduled run and a manual one read the same one. ~/wiki/INGEST.md:

Then the schedule, which is yours. On macOS a LaunchAgent is the supported mechanism and it survives the laptop being asleep at the appointed minute, where a cron job simply misses it. Save this as ~/Library/LaunchAgents/com.yourname.wiki-ingest.plist:

Load it once with launchctl bootstrap gui/$(id -u) and the path to that file. Or, if you already keep a crontab, one line instead:

Two things bite about that line. Cron runs with a minimal PATH, so the agent binary needs its absolute path or the job fails silently. And on current macOS a cron job that reads directories under ~/Documents or ~/Desktop needs Full Disk Access granted to /usr/sbin/cron in System Settings.

The join between a meeting and its write-up is the filename convention you chose, and nothing else: there is no field on a saved meeting recording where its notes landed. A year in, the themes pages answer the question nobody wants to ask directly, which is whether the answers given to the same question in March and in November were the same answer.

• No speaker labels. There is no diarisation, so nothing in the transcript says who said which sentence, and the minutes cannot attribute a line to a person from the audio. Attribution comes from what was said out loud ("Sam will take the migration"), from a connected calendar's attendee list, or from you typing the names. The **Present:** line is a line you fill in.
• macOS only. macOS 13 or later. The Windows audio path is designed and not released, and the scheduling half of this page is macOS-specific by construction.
• The agent is yours. It runs as your own CLI with your own key. Earkeep does not resell model access and there is no per-meeting AI fee, which is what makes it €39 once. What the agent does with the context after Earkeep hands it over is a property of the agent you chose, not of Earkeep. Setting up Claude Code covers installing and signing in to one of them.
• Nothing writes back. The MCP server is read-only, with exactly two tools, so a scheduled agent can read your meetings and cannot create or edit a meeting, a note or a file through it. The wiki is a one-way derivative of the record.

• Strategic planning meeting agenda template: where the plan being announced was actually decided, and where the rejected options are.
• Staff meeting agenda template: the smaller room where what the all-hands announced turns into work for each area.
• Telling people you're recording: what to say before capturing a meeting with a hundred people in it.
• All meeting minutes templates.