Quillm

Adds or updates rows in a dataset. This is how you "add new data" or "correct a number": every view reading the dataset shows the change immediately, with no view edit needed. Rows are matched on the dataset's key columns: a new key inserts, an existing key is merged (only the columns you send are overwritten). Safe to re-run. Values must match the column types (number columns take JSON numbers, not strings like "$1,200"; date columns take ISO strings like "2026-03-01" or "2026-03"). Unknown or not applicable is null, never "" or 0. It writes only to this Quillm workspace and calls no other service.

upsert_rows

Changes a dataset's schema or metadata: add columns, remove columns, fix a column's description or type (update_columns), change the description or update cadence. Keep descriptions true: when you start storing a new category or scenario, update the column description that lists them. Existing rows keep their values; new columns are null until you upsert values. The key cannot be changed; create a new dataset for that.

update_dataset

Creates a named table that views read with useDataset(name). Define typed columns and a `key`: the column(s) that uniquely identify a row (e.g. ["month"], or ["date","campaign"]). The key makes updates idempotent: pushing a row with an existing key updates it instead of duplicating it. You can pass initial `rows` in the same call. Prefer tidy, raw-ish data (one row per month/day/entity with numeric columns) over pre-aggregated text; views compute totals and deltas themselves. Check get_workspace first: if a dataset with this data already exists, add to it with upsert_rows.

create_dataset

Creates a new view: a live page at a stable URL in the user's dashboard app. `source` is ONE file of React (JSX/TSX) that default-exports a component and reads its data with useDataset("dataset_name") from "quillm". Do not embed data in the source; create the dataset first (create_dataset), then the view. Pass the `brief`: the user's request in their words, and what you asked them and what they answered. If the request does not say who reads the page, what for, and which number matters most, ask the user first; a brief that is too thin is refused once with the questions to ask. Decide the `questions` next: what the user asked for is the scope, and a small request gets a small page. The view is compiled, test-rendered in a real browser and reviewed immediately: the response tells you whether it rendered, lists errors, and lists what a reader would still be missing, so read it and fix problems with update_view. The checks use two outside services unless the workspace turned them off: TypeSafe reads the brief, and OpenAI reads a screenshot and the page's text. Call get_guide first for the template and available libraries. Check get_workspace first: if a suitable view already exists, use update_view instead.

create_view

Permanently deletes a dataset and all its rows. Refuses if any view still reads it, unless force=true. Only do this when the user asked for it.

delete_dataset

Permanently removes a view from the user's workspace. Only do this when the user asked for it.

delete_view

Deletes rows from a dataset by key, e.g. keys: [{"month": "2026-01"}], or every row with all=true (use before re-loading a dataset from scratch).

delete_rows

Saves durable notes for future agents (including yourself in a later session, with no memory of this one). Use it to record what is NOT obvious from the code: where the data comes from and how to refresh it (queries, API calls, file paths), definitions ("MRR excludes one-off invoices"), and user preferences ("Kev wants EUR, weekly granularity"). Notes attach to a view, a dataset, or the whole workspace, and are shown in get_workspace / get_view. Replaces the previous notes for that target, so include what should be kept.

write_notes

Shows the user a page inside the chat, live, with today's data, in apps that display pages (ChatGPT, Claude); elsewhere it returns the page's link. Call it once, last, when you have finished creating or changing a page or its data (after the review is clean), and when the user asks to see, open or show a page. create_view, update_view and upsert_rows show nothing in the chat themselves. Without a slug, lists every page.

open_page

Returns one view in full: metadata, agent notes, revision history, the schema of every dataset it reads, and its React source code. Always call this before update_view so your edits match the current source exactly.

get_view