We Did Not Invent Our All-Hands

We started running an all-hands this month. We did not design it. We read how somebody else runs one, took most of it unchanged, and spent our own thinking only on the handful of places where our situation is genuinely different from the one he was writing about.

The somebody is Gokul Rajaram, and this is a thank-you note with the receipts in it, because a thank-you without the specifics is just a compliment.

What we took, and did not improve on

Three acts. People, then alignment, then questions. Not four, not a rotating agenda, not a different shape each week. The shape is the same every Friday and that is the feature.

The 15/60/25 split. Fifteen percent of the meeting on people, sixty on alignment, twenty five on questions. In our thirty minutes that is five minutes, fifteen minutes, and nine, with a minute to close. Writing the split down as a proportion rather than as minutes is what makes it survive a meeting that runs long: when the clock hits 0:20 the scoreboard stops, finished or not, because the last quarter is not the part you borrow from.

People first, and first means first. The order of an agenda is a statement about what the company is for, made weekly, whether or not anybody intends it. Putting new joiners and named credit ahead of the numbers is a small thing that would be an enormous thing to reverse.

Rotate the presenter. This is the single line that made the whole framework adoptable here. An all-hands that one person owns is a part-time job for that person, and it eventually sounds like one voice reading a recording. Ours rotates through five people. The founder is on the rota like everyone else, and on the weeks that are not his he attends, answers questions, and does not run the meeting.

Two initiatives, not five. Peanut-buttering five initiatives across a thirty minute meeting is how nobody remembers any of them. Two, each with a named owner who presents it. A third only if something is genuinely on fire.

Attendance is the metric. No weekly survey. People vote with their feet, and a weekly survey is another job for somebody. Survey once a quarter if at all.

Feed people. Friday afternoon, standing between people and their weekend. Feed them.

Every one of those was adopted without argument, and they share a property worth naming: each is a rule that removes work rather than adding it. That is rarer in operating advice than it sounds.

The four things we changed, which are ours and not his advice

The pack generates itself, because we have no staff to curate it. Every all-hands framework we have seen assumes somebody has an hour a week to assemble content. Nobody here does and nobody here will. So scripts/ops/all-hands.mjs reads the systems that are already telling the truth, git history, the live MVP scorecard, the committed coverage measurements, the release timeline, and prints the whole deck as markdown in about a minute. Nobody writes a status report. Nobody is chased for an update. The presenter fills exactly two blanks: the Why, in their own words, and the two initiatives.

The scoreboard is shaped like the strategy, so the agenda restates it and the founder does not have to. Gokul is right that you can never restate strategy often enough, and right that a founder saying it every week is exhausting and starts to sound like a recording. Our answer was to stop saying it. The numbers section is always the same four blocks in the same order, and those four blocks are the four moats we believe are the only durable ones: brand, product, customer service, and trust and storytelling. Twenty six Fridays of seeing those four headings in that order teaches the strategy better than twenty six speeches about it. And when somebody cannot find where their work belongs in those four, that is a real signal rather than a formatting problem.

A number with no system behind it prints as "not measured yet". The tempting version of a generated pack leaves a blank for the important numbers and has somebody type them in weekly, and that is exactly how a pack becomes a status report and a status report becomes a job. So a missing number is printed as missing. Right now the pack says, every single week, that removal requests received and honoured are not measured yet. We promise removal publicly, on every listing, to people who never asked to be in our directory. Seeing that gap in the same place at the same time every Friday is uncomfortable in the way it should be.

The human and agent commit split prints every week. Of the 1,477 commits in this repository, 649 were written by an AI agent. Forty four percent. In the section of the meeting that exists specifically to credit people, a single line reading "the team shipped this many changes this week" would be a lie told by omission, in the one act where a lie would do the most damage. So the generator counts them separately and prints both numbers, forever, next to the names of the people who wrote the rest.

Where his advice had to bend, and why that is on us

Here is the honest note. Gokul was writing for companies that have people to run this. We are very small, and we have decided not to hire an operations team at all. Instead we are building an agent to do the chief-of-staff job, and its definition now lives in this repository next to the code.

It runs the generator on Thursday. It chases the presenter for the two blanks, once, and if they are still empty on Friday morning the pack goes in with the blanks visible, because a gap everybody can see is self correcting and a gap quietly filled in by a machine is not. It collects and ranks the pre-asked questions. It runs the scorecard and reports the delta. And it keeps the ledger, which is the highest value thing it does: every "I will come back to you by Wednesday" said out loud in an all-hands becomes a tracked row, and the agent surfaces the overdue ones. That promise is the one most likely to evaporate, because the person who made it was standing up, answering fast, and completely meant it at the time.

What the agent explicitly does not do is decide strategy, write the Why, choose the initiatives, or present. Those are human and they stay human. An agent that drafts the Why produces something in nobody's voice, which is worse than a blank.

None of that is what Gokul was writing about. Advice written for a company with an operations function, applied by a company that has decided never to have one, is going to bend somewhere. It bent here. If it breaks, that is ours, not his.

The thank-you, stated plainly

We did not have to spend a single week figuring out the shape of this meeting, and we would have got the shape wrong. Somebody with a great deal more experience than us wrote down what he had learned, published it for free, asked nothing, and we picked it up and ran it. That is a real gift and it is worth saying so in public rather than in a private note nobody can check.

Gokul Rajaram holds no role at 🤫 hussh, HushOne, Inc., or Hushh Technologies Corporation. We have no affiliation with him, and none with Square, Block, or any other company he has worked for, advised, or served as a director of. He has not reviewed this post, has no involvement in our work, and did not write his article for us. Naming him shares our admiration and what his public work has given the world; it implies no affiliation with, endorsement by, or sponsorship by him or by any organisation he is associated with.

The framework as we run it is in the open at docs/ops/all-hands.md, and the generator that builds the pack is at scripts/ops/all-hands.mjs. Both are short. Take either, the way we took his.