The weekly leadership deck is not hard work. It is careful work, done under time pressure, by somebody who could be doing something else — and the two failures it produces are both quiet.
The first is a stale number. Somebody could not reach a figure on Monday morning, so last week's went in, or a zero did, and the meeting discussed it as though it were this week's. Nobody notices, because a wrong number looks exactly like a right one. The second is a deck of everything: forty rows in no particular order, read out in sequence, so the hour goes on reading rather than deciding and the one metric that needed twenty minutes gets ninety seconds at the end.
So this is built around both. Every figure is computed in code before any sentence exists — week on week, the four-week average, the distance from plan, the status of each metric, and the order the rows appear in. The step that writes the prose has no tools at all: it cannot reach a record, cannot fetch a number, and cannot recompute one, because everything it is allowed to say is already in front of it, already worked out and already formatted as the string it must copy. A number that is not in that input cannot be written, and the checking step rejects the whole deck rather than publish one that was.
An empty cell gets the same care, which sounds like a small thing until you hear a deck read out. There are five genuinely different reasons a figure can be missing and the deck says which: the record exists and holds no value this week; the source keeps no weekly history at all, which is true of every goal metric because a goal stores one reading rather than a series; nobody set a target; there were fewer than three underlying records, so a percentage would be noise with a percent sign on it; or the baseline was zero, so a ratio has no meaning. A dash covers all five, and in a room a dash is heard as "flat".
What earns a spoken position is the other half of the design. Every metric is in the table — a metric missing from the table is the failure this replaces — but only four things are said out loud. One metric opens the meeting: the one furthest behind its plan in the direction that matters for it. Four tiles go to the misses before anything else, because the thing a room must hear is not the thing whoever set the deck up likes to show. One bullet each goes to the metrics that moved materially or are behind plan. And the asks name one person each — but only where nobody has accounted for the miss yet: a goal that is behind plan whose owner posted a check-in inside their own agreed cadence has already answered the question, and asking it again every week is how a "waiting on you" list becomes a list nobody reads.
The charts are inside the document rather than beside it, and they are built in code and copied. That matters more than it sounds: a chart block whose id is written out by hand parses as malformed and renders as an empty box, with no error in the run, the document or any log. So the blocks are produced by the same serialiser a person's own edits go through, they are handed to the writer finished, the checking step requires every one in the deck to be unchanged, and the deck is parsed again after it is written — a block that no longer draws is a named failure rather than a blank rectangle somebody assumes is loading.
Then there is what it declines to do. It reads your goals, your board and your agent runs, and nothing else, because every figure in the deck has to be recomputable from a record this workspace holds. A number pulled from a CRM by a model in a tool loop would sit in the same table, the same typeface and the same citation as one the run derived itself, and nothing downstream could check it — which in a document that is read aloud launders it. It also files nothing by default. A deck whose side effect is five new tickets assigned to the people sitting in the room, created before anybody has spoken, is a deck people learn to distrust; it names who is needed, and the room decides. If you want the chasing on your own board, you can turn filing on: up to five tickets, addressed to you rather than to the people in the room, chosen and titled before the run starts rather than by the model, skipped when last week's ticket for that ask is still open, and every ask over the limit named in the deck itself so nothing disappears into a cap. It can create those tickets and nothing else — there is no step in it that can close, move or reassign one.
And it says what it could not see, under its own heading. Tickets belonging to a team the deck cannot see are dropped entirely rather than shown redacted — and counted, so the deck states how many, and says plainly that every ticket figure in it is a minimum. That is the one thing a summary must never get wrong: a throughput number that is quietly smaller than the truth is still a number somebody makes a decision on.