The Problem It Solves
A curator reads your inbox and finds eleven things that might need you. The naive design asks you eleven questions. You answer three, the rest scroll away, and next morning it asks again. Within a week the brief is a wall of unanswered questions and you have stopped reading it.
The alternative: the brief is the surface layer of an incremental-handling system. Items do not dump on you all at once; each brief surfaces a small slice, and you dispose of items one at a time, over the day, in any session, from any device.
Pre-drafted Tasks
Every actionable item a curator surfaces becomes a task with the unconfirmed label, at low priority, before the brief is written. Its description carries provenance — which curator, which date, the source type, and a deep-link to the thread when one exists:
<!-- src: email curator daily digest 2026-05-07 -->
<!-- src-type: email -->
<!-- src-url: https://…/thread/… -->
A de-duplication pass runs first: if an open unconfirmed task already matches the title fingerprint, the new sighting appends a “still surfacing” note instead of creating a twin. (The de-dup pass was skipped once in a chat session and produced a duplicate; the feedback entry made it mandatory in every path.)
Priority defaults to 3. Curators may write priority 2 directly only for a narrow carve-out: a deadline within seven days, past-due billing, a security alert, a conflict on today’s calendar, or an external decision pending you. Everything else waits in the backlog until you confirm it.
One-Word Dispositions
The brief shows each item framed as its decision. You reply with a short instruction, and the chief of staff executes it:
| You say | It does |
|---|---|
| confirm / raise to P1 | drops unconfirmed, sets the priority |
| dismiss / close | closes with reason “dismissed from the brief” |
| done | closes with reason “already done” |
| defer 7d | adds a follow-up date and the waiting label |
| re-label to venture-y | swaps the venture label |
| task it friday | pushes it as a dated, completable task into your task app |
| block it tomorrow 9–11 | books a work-block on the calendar pointing at the task |
Voice is the primary path. You record one note hitting items in the order they appear in the brief — “for the subscription one, cancel; for the invite, decline; the lawyer thing, defer a week” — and the assistant parses by position-in-section, never by task ID. The typed shorthand (close b50) still works on the laptop, but you should never have to type an ID from a phone.
Time-to-Live Rules
Some items you will never answer, and silence carries information:
- Unconfirmed items older than seven days appear in the weekly review’s drift section, batched, for a one-pass decision.
- “Was this you?” security alerts — new device, new sign-in, verification code — are auto-closed after fourteen days at ordinary priority. If it had not been you, you would have said so. High-priority ones are exempt, because a raised priority means the assistant had a specific reason to doubt. Closures are listed in the brief’s Atlas-changes line, so a wrong one is one reopen away.
- Confirmation-gated items are not evidence. A payable that is still open because nobody confirmed payment is a confirmation request, not a delinquency. The line says “still open on my side — did you pay this?” and never “12 days past due.”
Why It Works
Three properties, each a deliberate choice:
- The cost asymmetry favours surfacing. Dismissing costs you five seconds; a missed signal can cost weeks. So the system over-surfaces by default, and a single dismissal never becomes a suppression rule. Only an explicit “stop surfacing X” suppresses a class.
- Nothing depends on you answering today. The task exists; the brief is just where it is visible. Answer in a week, from the review surface, in bulk — the graph is the same.
- Provenance survives disposition. When you say “close it” the reason is recorded; when you ask “why did this exist?” the source link is in the description.
The Laptop Side
For heavy days, the same dispositions are buttons: the review surface renders the brief with every task reference clickable, opens a modal with the full task and the disposition menu, and walks the stale buckets one at a time. See The review surface.
Hand this to your chief of staff. Plain Markdown: /guides/incremental-handling.md. Tell it: "Read this guide and tell me what in our system it would change."
Also for your Claude: whole-site index · bootstrap prompt · guide feed