A summary should tell you what moved.

A while back we wrote about why timeline labels are navigation, not summaries. Every row in your history is a scan-first headline for one thinking event. It answers two questions: what happened here, and do I care enough to open it.

Then we shipped the obvious next thing. Pick a range on the timeline, hit Summarize, get a short note about the period. The first version worked on the first try, read competently, and said almost nothing:

A day focused on provenance, voice memos, and positioning. The team worked to solidify provenance with edits to FEATURE-5, FEATURE-6, and FEATURE-7. Voice memo handling was enhanced. Several notes addressed positioning.

Accurate. Useless. It is a table of contents grouped by theme. It tells you the topics that were in the air. The person reading a week of their own work is not asking about the topics.

This is the exact mistake from the row post, lifted to the aggregate. There, the trap was the topic summary: “Notes about pricing.” Here, the trap is the topic re-grouping: “A week of pricing and provenance work, touched PRICING-1, FEATURE-5.” Same failure, bigger surface.

And it is worse at this level, for a reason that took us a second to see. The summary is built from rows that are already excellent one-line headlines. If the summary just reshuffles them by topic, it hands you less than scrolling the list would. You lose the specific headlines and gain a vague theme. The summary has to earn its place by saying something no single row can.

A row and a window summary are not the same object at different sizes. They answer different questions.

1. A row asks: what happened here, and should I open it?
2. A window asks: where do things stand, and how did they get here?

The row is a waypoint. The window is the map of the journey. Strip away the software and it is the most ordinary human instinct: the weekly review, the “catch me up,” the previously on recap before the next episode. What a person wants, in order, is the shape of the period, the threads that actually moved, and the live edge of where the work is now pointed.

Here is the leverage we almost missed. Every line feeding the summary is already a curated headline that a row generator distilled to a verb, an object, and a twist. The summary is not compressing raw content. It is synthesizing across signal that is already clean.

That changes the job from re-describe to connect. The value the summary adds is the line no row can draw on its own:

Resolved the pricing question: opened it Monday, ruled out usage-based midweek, committed to per-seat Friday.

Three rows on three days become one thread with a beginning, a turn, and an end. That is the thing you cannot get by reading the rows one at a time, and it is the only reason the summary deserves to exist.

A topic is “provenance.” A thread is “hardened provenance from an idea into a committed core feature and locked it at a milestone.” A topic is a noun you could file under. A thread has movement and, ideally, a place it landed.

So every bullet leads with a verb of movement, the same vocabulary the row labels use. Resolved, Reframed, Reversed, Committed, Narrowed, Opened, Explored, Hardened. The verb tells you what kind of thinking happened before you read another word. The worst phrasings are the ones that feel like work but carry no signal: worked on, made edits to, focused on. They are the aggregate version of “Created,” and we ban them for the same reason.

The first version of the prompt adjusted one thing by range: word count. That was the wrong dial. The range should change the question, not just the length of the answer.

• A day or two. Where did I leave off, and what is still in motion? Lead with the live edge. This is a resume cue, not a retrospective.
• A week. What actually moved? Collapse the events into the few threads that advanced; say which resolved and which are still open.
• A month or a quarter. What was the arc? Name the big shifts, the reversals, the abandoned directions, and the throughline. Individual notes matter less than the movement between them.

That last one is the real bar from the row post, scaled up: let a person replay the evolution of their thinking without reopening every artifact.

A summary that names FEATURE-5 and leaves it as dead text is a strange kind of dead end. It points at the source and gives you no way to get there. So the last move was to close the loop.

Every display id the summary cites becomes a live link back to that note. Click it and the note opens on the canvas, exactly like any mention anywhere else in the product. The summary is no longer just a thing you read. It is itself navigation, back into the history it describes.

The mechanics lean on a decision we had already made for the editor and the CLI. There is one mention syntax, [FEATURE-5: Title](relationship:references), and a save path that already turns it into a real linked node. So we did not ask the model to produce that finicky shape, where it would paraphrase the title or invent an id. The model writes plain ids. A single platform-owned pass rewrites each real id, using the authoritative title from the events themselves. An id that does not exist in the window stays inert text. Same principle as the labels: the platform owns the link, so it always works.

A window summary is movement, not inventory. It is the story the rows add up to, not the rows restated.

A summary that lists what you touched fails. A summary that regroups it by topic fails. A summary that tells you a question got resolved, a direction reversed, an idea hardened into a decision, and lets you click straight back into any of it, passes.

Summarize the movement. Everything else is a table of contents.