The pantry

Everything in a library sits on a shelf. You name it, say in a line what it is for, and keep longer notes on the shelf itself. A shelf lists its documents in natural title order, so “Week 2” comes before “Week 10”, and it has an Index to this shelf of its own.

Shelves do not nest. A document lives on one shelf at a time; Move puts it on another.

A document has a title, an overview you write like the opening of a chapter, and as many numbered findings as it needs. Each document gets a stable number, No. 12, that never changes and is never reused, and a facts line that counts its findings, words, reading time and highlights.

A document’s actions sit in one row: Star, Focus, Share, Move and Edit. Your assistant can create documents for you by the dozen; Edit opens one in the editor.

A finding is a full page of its own: a title, a body with headings, lists, quotations, tables and code, a list of source links, and one of ten types — Observation, Evidence, Insight, Question, Decision, Suggestion, Problem, Risk, Opportunity or Bug. Mark the ones that matter most as Key.

Findings are numbered by their order, 3 of 8, and you can drag them into a new order. Each has its own address, a previous and next, and an Indexed under line listing the index terms that point to it. Findings are not starred; documents are.

Your assistant does the heavy lifting; the editor is where the work becomes yours. Press Edit on a document or finding and write straight into the page: headings, lists, quotations, tables, code blocks and links, with Markdown pasted in turned into the same. Drag findings into a new order, change a finding’s type, mark the ones that matter as Key, add or remove sources, delete what does not belong.

Saves happen a second after you stop typing, and the page tells you in words when they have. If your assistant changes a document while you are editing it, know.sh says so and lets you keep your version or take theirs, and every earlier version stays in Revisions.

know.sh reads the titles and opening lines of your findings and picks out the words a back-of-book index would: names of people and places, titles, organisations, acronyms and technical terms — Naomi, Philippians, NDVI, useLoader — once they recur across two documents or three findings. Each is followed by the numbers of the findings where it appears; plurals get a See, terms that keep company get a See also, and the number of a finding you have highlighted is set in bold.

The index is made of single words that recur, so it is a cast list for a novel, a table of cases for a law outline and a glossary of acronyms for a course. Everyday words and phrases, such as buttermilk or pot roast, are left to Look up, which searches every word. Each shelf and each document has its own slice of the index.

Ask your assistant for a quiz on a document and it writes five to ten questions with know.sh’s quiz tools: multiple choice, select all that apply, true or false, or a typed answer that must match exactly, capitals and spaces included. Every question cites the findings it was drawn from, so a wrong answer sends you back to the right page. You take the quiz in know.sh, on the web or the iOS app.

When you finish, know.sh schedules the quiz to come back under Revisit: in three days if you scored under 60 per cent, a week if you scored 60 or more, three weeks at 80 or more. It is a simple rule, not a flashcard system, and it is enough to make you come back.

Select text while reading and choose Highlight. The highlight stays attached to its passage when the text around it is edited, and if the passage itself is rewritten you are asked to re-attach it.

Highlights feed the rest of the library: highlighted findings are bold in the index, Open at on the title page takes you back to your latest one, and Revisit points you back to a document whose highlights you have left alone for a month.

Select a passage and choose Note; the note sits in the margin beside it on wide screens and under it on narrow ones. Notes can have replies, one level deep. There is no resolve button: a margin note is meant to stay, like a pencilled note in a book.

Notes are yours and your assistant’s. Readers of a public link do not see them.

When your assistant notices something you did not ask about — a date that disagrees with an earlier finding, a claim without a source — it can leave a proposed note: a margin note with a dashed rule and Accept and Dismiss under it. Accepted notes become ordinary notes; dismissed ones go away. Waiting proposals are listed under Revisit on the title page.

Proposals are notes, not edits: nothing in your text changes through one.

Each document and finding keeps its revisions: when, who — you, or an assistant connected through MCP — and what changed. Restore puts an earlier version back.

Revisions make it safe to let an assistant loose on a document, and they keep last term’s syllabus or last season’s plan without a pile of copies.

Share makes a public, read-only link to one document: its title, overview and findings, without your highlights or notes. The link is shown once, so copy it then. You can set a password, set an expiry date, leave individual findings out from each finding’s page, replace the link so the old one stops working, or turn it off. The panel tells you how many times it has been read.

Links are per document. Shelves, whole libraries and quizzes are not shared, and readers cannot comment or edit. Anything private — health, children, clients, money, door codes — stays off public links.

Look up (⌘K, or / anywhere) searches as you type: index terms, documents, findings, highlights and shelves, with full-text search behind it, so the everyday words the index leaves out — an ingredient, a symptom, a phrase — are found here.

Star marks a document with ★ wherever it appears and gathers starred documents in Contents. Focus puts a document in the spotlight for seven days: it rises on the title page, and when the week ends Revisit tells you, so an unfinished chapter or an exam unit does not quietly slip away.

Contents lays the library out like the front of a book, shelf by shelf with each document’s number, or month by month in the order you filed them. Views for Recent, Starred and Shared sit beside them, with a filter.

A campaign turns a body of research into work to do. It holds operation sets, and each set holds operations — the steps — with a status (draft, ready, active, blocked, done) and dependencies: one operation blocks another, informs it, or is related. A graph view lays the plan out so you can see what is waiting on what.

Campaigns suit the projects around a book or course: a launch, a term’s schedule, a restoration. They are planning for one person; there are no assignees. An assistant connected with the planning permissions, which the Claude Code plugin asks for today, can lay a plan out for you from your research.

know.sh does not ask you to adopt its AI. It runs an MCP server, the open standard assistants use to reach tools, so the assistant you already use becomes the one that works in your library. You sign in once with your know.sh account; there is no key to copy. The server’s address is on the support page.

• Claude, on the web, desktop or phone: Customize → Connectors → Add custom connector, paste the address and sign in. Every Claude plan can add one; on Team and Enterprise an owner adds it first. In Claude Code, install the know.sh plugin or add the address as an HTTP server, then run /mcp to sign in.
• ChatGPT, on Plus, Pro, Business, Enterprise or Edu on the web: turn on Developer mode, add know.sh as an app with OAuth, and ChatGPT asks before it changes anything. Sign-in from ChatGPT is still rolling out; the support page says when it is ready.
• A local model: run it in Ollama and connect know.sh from a chat app that speaks MCP and signs in with OAuth, such as Open WebUI, Goose or VS Code in agent mode. Choose a model that supports tool calling and give it a long context (64k or more).

Once connected, your assistant can search the library, read documents and findings, create and reorganise them by the dozen, add notes and highlights, propose notes for you to accept, and write quizzes. If it can search the web, it can research a question and file what it finds, with sources. Everything it writes is marked in the revision history as coming from an assistant, so you can see and undo it. Switch assistants whenever you like; the library stays the same.

Local models work, with limits: know.sh offers dozens of tools, and smaller models pick the wrong one or skip steps more often than hosted ones. For long, many-step research, a hosted assistant is more reliable.

The leading assistants can run a prompt on a schedule. Claude runs scheduled tasks on its Pro, Max, Team and Enterprise plans, hourly, daily or weekly, in the cloud, even while your computer is asleep. ChatGPT offers scheduled tasks on every plan: daily on the free plan, hourly on paid ones, with the results arriving by notification or email. Both can search the web during a run.

A scheduled run is the way to keep a library current: this week’s injury news for your fantasy league, the AI models released since Monday, new papers on your thesis topic. When the result arrives, reply file this in know.sh and your assistant adds it as findings with their sources. Some setups can file straight into know.sh during the scheduled run itself; this is newer and not every plan or connector supports it yet, so check the revision history after each run either way.

A link inside the library — from a finding to another document, or from a campaign step to the research behind it — shows a small mark and, on hover, a preview of what it points to. Your assistant writes these links when you ask it to connect related material; you can also paste a link to a document or finding into the text.

The iOS app reads and edits the same library, with the thumb index under your thumb: read a document at the stove or in the pew, highlight and note, and take a quiz on the bus. It is in beta on TestFlight today.