AI

What the AI does, and the one thing it never does.

Six features, built on one rule. Filing what you already decided is clerical, and a model may do it at any moment. Deciding what your vague thing breaks into is the descent, and that stays yours. So every feature here ends the same way: a set of changes you could have made by hand, shown line by line, waiting for you to edit and commit. Nothing is written by a model.

Chat with a note

Talk it through. Then resolve.

Open a chat under any note when you're stuck. It's disposable: nothing you say there goes on the record, which is what makes it cheap to think out loud. It has read the note and what's around it, can search the rest of your notes, and can look things up on the web. When the talk gets somewhere, press Resolve. The conclusion is proposed as a comment on the thread, and anything it turned up as draft notes beneath, each yours to edit or strike before you commit.

Why do I feel constantly behind? Think

  1. I keep saying I'll do the course on weekends and I never do.

  2. Three weekends in a row, going by your thread. What actually happened on those Saturday mornings?

  3. Honestly, email. Then it's lunchtime.

  4. Then the course isn't losing to tiredness, it's losing to the inbox. Worth blocking the morning before email, and checking how much of that inbox is newsletters?

Resolution

From this chat

  • +
    The course loses to the inbox, not to tiredness. Try the morning before email.

    comment on the thread

  • +
    Action Block Saturday morning for the course backlog

    new, under this note

Smart add

Paste the brain‑dump. Get the notes.

When what's in the capture box runs past a line or two, a second button appears: Sort this out. One call turns the text into notes, lines split or merged, a kind on each, topics from your @mentions, dates read out of the words. It checks against what you already have, so “book the dentist” doesn't become a second note. The result waits in the same panel as everything else.

renew the car registration before friday, book the dentist, the gutter thing @house - get two quotes, and work out whether we actually want the pergola or just some shade

Sorting this out

4 lines, 3 new notes

  • +
    Action Renew the car registration

    due Fri 24 Oct

  • +
    Action Book the dentist

    dropped: already a note, open since Tuesday

  • +
    Action Get two quotes for the gutters

    @house

  • +
    Think Do we want the pergola, or just some shade?

    @house

Review

The app does the arithmetic. The model does the meaning.

Review's queue is ranked by things only the app can see: how long since you looked at something, how often you've opened it without working on it, what's due, what you've put off. Then a model reads the shortlist, because a weighted sum can't tell that a daily page is a container rather than a task, or that “inflate the pools” needs you standing next to the pools. It picks five to seven, says in a few words why now, adds a nudge when there's something worth saying, and guesses how long. Light looks for things to finish. Focus looks for things to open up. Say you're on a train and it skips what needs a desk.

  • Get two quotes for the gutters Action

    Coming up · ~5m

    ✦ Two emails. The second quote is the one you keep not sending.

  • How do other people deal with this? Research

    Quick, anywhere · ~15m

    ✦ One search on the train would close it.

  • Why do I feel constantly behind? Think

    Waiting 12 days · ~10m

    ✦ Is the sharper thought underneath it still the right one?

The stand-up

A conversation over everything. Not a form.

Whenever you like, not weekly, open a stand-up. It starts with one open question, picked by the app rather than generated: what's taking your attention at the moment? Or, when it remembers, how's the extension going? You talk. Every follow-up comes from a gap between what you just said and what your notes hold: a note still open that you've just called done, a thing you mentioned that isn't captured anywhere. Four follow-ups, then it offers to close, and you can keep going. The close proposes captures as new notes, corrections to stale ones, what a topic is for, and an edit to the notebook. The transcript lands on the stand-up's thread as it was said.

Stand-up

it's been 11 days

  1. How's the extension going?

  2. Signed off on Tuesday, finally. Now it's the bathroom quote I keep not chasing.

  3. “Get the extension signed off” is still open here, and there's nothing about a bathroom quote. Has a quote come in, or is chasing it the job?

  4. Chasing it. Two builders, neither has replied.

Close

From this stand-up

  • ✓
    Action Get the extension signed off

    open → done

  • +
    Action Chase the two builders about the bathroom quote

    @house

  • ~
    Notebook · Working on

    “Getting the extension signed off” → “The bathroom, now the extension is done”

The notebook

What it knows about you, in a note you can edit.

The memory is one note under four headings: working on, what matters, about them, don't raise. Every chat and every connected app reads it first, so none of them start from nothing. It is written the way everything else is: proposed at the close as an edit against the current text, never silently. If it drifts wrong, change it like any note. If it gets long, the close is asked to shorten rather than append.

Notebook Scratch

edited at the last stand-up · you can edit it too

Working on
The bathroom, now the extension is done.
The course backlog, Saturdays.
What matters
A house we actually want to live in.
Not losing the piano.
About them
Evenings are when work spills over.
Doesn't finish long books.
Don't raise
@admin. A chore bucket, not a subject.

Connected apps

Bring your own assistant.

Connect Claude, or anything else that speaks MCP, from Settings. It signs in as you through an ordinary consent screen, and from then on it can read your notes the way the built-in chat does and propose changes the same way. Nothing it proposes is applied: it lands under Proposals with its summary and your request quoted beside it, and you commit what you keep. Disconnect any app in one click. The server behind it calls no model of its own, so it spends none of your allowance.

get_context
The date where you are, your notebook, and your topics with their goals. Called first.
search_notes
Find notes by words in the title, description or thread.
list_notes
Every note in a topic, of a kind, or in a state. “What is open in @garden.”
read_note
One note in full: what it sits under, its parts, its thread.
topic_overview
What a topic holds and how much of it is done. Not what it is for; only you can say that.
read_board
The board as it’s drawn: each planned note by topic and column, and what it waits on.
propose_changes
A set of changes, with your request quoted beside it. Waits under Proposals.
list_proposals
What it and other apps have proposed, and what became of each.

Proposal

From Claude · 3 changes

Waiting for you
  • +
    Action Email the course provider about the refund

    new, under Why do I feel constantly behind? · @courses

  • ~
    Think Why do I feel constantly behind?

    add topic @energy

  • +

    comment on Things that might be stopping me getting ahead

    Sleep came up in three notes this month. Is that one problem rather than three?

The rules

What every feature keeps to.

  • Nothing is written by a model.

    Every result is a set of changes you could have made by hand: a note, a comment, a tick, a tag, a date, a link, a move, an edit. Each shows what it changes from and to. Remove any line, edit any text, then commit.

  • Every change carries what it saw.

    If a note has moved on since the model read it, that line is dropped at commit and you are told. A stale suggestion can never overwrite something you typed.

  • It can rearrange. It can’t delete.

    It can propose moving a note under another, taking a link away, or saying what waits on what, each as a line you can strike. It can never delete anything: the furthest it goes is archiving a note.

  • Questions, not interpretations.

    The stand-up asks about the work and the record. It never says “it sounds like”. Review gives a nudge when there is something worth saying, and nothing when there isn’t.

  • It reads as you.

    Every read and write runs under your own account, the connected apps’ included. Nothing it can see is anything you couldn’t.

  • A daily allowance. Nothing unattended.

    Model calls are counted per day, and past the line the app says so. No model runs on a schedule or without you in front of it.

Write the vague thing down.

Then come back to it when you have a moment. That's the whole method.

Get started