Quill Writing Assistant

Quill is the writing assistant built into WorkBOS. It watches every prose text box in the product - notes, templates, sequences, emails, tickets, descriptions, anywhere you write - and it watches the rich editors too: Drive documents, including while somebody else is typing in the same document with you, and the document templates you send to clients. It underlines spelling slips, grammar slips and words your workspace has banned, with one-tap fixes on each. Identifier fields (emails, URLs, codes) and code editors are deliberately left alone. The underlines are computed entirely in your browser, from a dictionary and rules that ship with the app, and are sent nowhere. One part of the live layer is not like that, and it is named here rather than buried: whole-sentence correction sends each finished sentence to the AI provider to be checked, and it is ON unless somebody switches it off. Everything else waits until you explicitly ask Quill for an AI action. The Quill page (Marketing group) is its home, and a small Quill button sits under every supported field.

• Watching as you type: spelling against a built-in dictionary plus the workspace shared list and your own, with suggestions ordered by which word people actually use (so "wer" offers "were", not "war"), grammar and punctuation slips (its and it is, then and than, a and an, could of, doubled words, spacing), and your workspace banned or preferred terms. Finish a flagged word and the suggestion pops up right at the word: press Tab to take the top fix, Esc to dismiss, or hover or click any underline to reopen it, with one-tap fixes, Ignore, and Add to dictionary. Without a mouse: Alt and the down arrow (or Ctrl and space) open the card and put you on the first suggestion, Tab and Shift-Tab move through the rest of its buttons, and Esc closes it and puts your cursor back exactly where it was. Until v1.252.0 that was not possible at all: Tab was the only key that reached the card and it was wired to accept the first fix, so a keyboard user could take whatever Quill offered first and could not choose a different one, Ignore it, or add the word to the dictionary. Joined-up words get the split you meant (iwant becomes I want). The eye button under the field pauses or resumes watching for you (one personal setting, all fields); admins set the workspace default. The built-in dictionary is English: on text in another language Quill stays quiet rather than mis-marking it.
• What the grammar rules cover: subject-verb agreement, including working out the real subject when a phrase sits between it and the verb ("the list of outstanding items are on page two"), plural subjects with singular verbs ("the invoices was paid on Friday"), "each of the invoices are overdue", counting ("three box of A4 paper"), words that have no plural ("those information"), the possessive apostrophe ("next weeks delivery"), a tense that argues with the sentence’s own time words ("last Tuesday I send the signed contract"), word order in a question inside another sentence ("please let me know when will the parts arrive"), pronoun case ("between he and the landlord", "send the terms to Jane and I"), two finite verbs with a pronoun between them ("Were you have the file?"), comma splices and run-on sentences, confused pairs (its and it is, then and than, there and their, whose and who is, where and were, a and an, could of), a past tense where the participle belongs ("the report was ran twice"), doubled words and spacing. Eight rule packs carry all of that, fetched in the background when Quill starts rather than shipped inside every page you open. It is a fixed list rather than a language model: it catches what it has been taught and nothing else, which is exactly what the whole-sentence pass below is for. Guards run the other way too, so "a number of customers have asked", "the invoices that we sent are attached", "if I were you", "ten miles" and a colleague’s name are all left alone.
• Whole-sentence corrections: the underlines find the mistakes Quill has been taught to look for, which is several hundred patterns and still a fixed list. This finds the rest. Quill reads each finished sentence and offers the corrected version, the way a person editing your draft would, so a sentence that is wrong in a way no single word explains still gets fixed. It works in every text box and inside Drive documents. IT IS ON UNLESS SOMEBODY SWITCHES IT OFF, and it is the one part of the live layer that sends your text: each finished sentence goes to the AI provider connected to the platform to be checked. It shipped switched off, which meant nobody ever saw it, so the default changed and the disclosure moved onto the Quill page where somebody will actually read it. An admin can switch it off for the whole workspace, anyone can refuse it for themselves on the Quill page, the underlines stay on your device either way, and every sentence check is recorded in usage like any other AI action. It spends its own allowance (an hourly one per person), not the daily AI actions, so an hour of typing never uses up the buttons; the Studio tab shows both counts.
• One word wrong gets you both fixes. When a sentence has exactly one mistake in it, the word fix stays under the word - instant, exact and computed on your device - and the whole corrected sentence sits on the same card under "Or use the whole sentence", one tap away. Nothing is applied until you press it. When two or more things are wrong in the same sentence, the sentence offer takes over the underline instead, because accepting three corrections one at a time and still being left with the grammar helps nobody. A word Quill has never seen (a client’s surname, a product code) holds the sentence offer back until you add it to the dictionary, so a name is never rewritten into somebody else’s.
• Quill checks its own answer before it shows it to you. The corrected sentence goes back through the same grammar rules that found the mistake, and a rewrite that still trips one of them is refused and never offered. "War you has the manager’s report?" used to come back as "Were you has the manager’s report?" on a button saying the sentence had been put right; a wrong answer under a confident label is worse than no answer, so that one is now refused instead.
• How well it works, measured rather than asserted. On eighty sentences written by somebody who was not allowed to read a single rule, Quill catches 78% of the mistakes. On the set the rules were tuned against it catches everything, and that number means much less, which is why we do not quote it. On prose that was already correct it paints 6.73 marks per 1,000 words, down from 19.39 once passive voice stopped being on by default, and more than two thirds of paragraphs now draw no mark at all. For spelling, the word you meant is the first suggestion 86.8% of the time and is among the five shown 97.2% of the time. AND HERE IS THE NUMBER TO JUDGE US BY. The eighty-sentence set has itself been worked against now, so a third set was written blind on 1 September, a hundred sentences nobody had shown the engine. Quill catches 3.3% of the mistakes in it, and flags none of the forty correct sentences. Three point three percent is what this engine does on English it has never seen, the 78% and the 100% are what it does on English it was tuned against, and we would rather print the small number than let you find it yourself.
• Dictionary: two lists. The workspace list holds the words everybody uses, your product names, your clients, the words your trade uses, so nobody has to teach Quill the same word fifteen times; owners and admins look after it. Your own list is private to you. Add to dictionary on any suggestion adds to your own list.
• Kinds of suggestion: spelling, grammar and punctuation are always on, because a misspelling is not a matter of taste. Clarity (wordiness, redundancy), engagement (a word used too often, worn business phrases) and delivery (hedging, shouting, words with a neutral alternative) can each be switched off, by an admin for the whole workspace or by you for yourself, on the Quill page.
• Passive voice is off by default, and that is a measurement rather than an opinion. On 33,880 words of business writing that was already correct, Quill drew 657 underlines and 429 of them were passive voice: 65% of every mark the product made, on the one rule with no button to press. On forty correct business sentences the engine had never seen it wrongly flagged nine, including "we ask that the deposit be held in a client account", which is a subjunctive and not passive voice at all. The rule is unchanged and one checkbox on the Quill page turns it back on. If your workspace has switched the whole clarity family off, that still wins.
• Ignore remembers. Press Ignore on a suggestion and Quill stops raising it, and it stays that way after you reload. Those choices are kept in your own browser rather than on our servers, which is why they do not follow you to a different computer. The Quill page shows how many you have dismissed and can bring them all back.
• What Quill learns from you, exactly, because "self-learning" is the easiest claim in the world to make and this one is meant to be checked. Four things and no others. First, which suggestions people take and which they dismiss: every accept, ignore or add to dictionary is recorded as the kind of suggestion it was and what was done with it, and nothing about what you typed is stored, not the word, not the sentence, not the fix; the record cannot hold text. Admins see accepted against ignored per kind and per rule on the Usage tab. Second, which kinds of suggestion you keep ignoring: ignore the same kind five times in a row without taking one and Quill stops offering it to you, says so, and lists it under My settings where one click brings it back; taking one resets the count; this lives in your browser and is sent nowhere; spelling, grammar and punctuation can never be stopped this way, for the same reason the workspace cannot switch them off. Third, corrections your workspace has accepted: when somebody takes a whole-sentence correction and exactly one word changed, Quill remembers that pair of two words for the workspace, and once it has been accepted twice everyone gets it as an instant on-device fix, when the misspelt word is not one the dictionary knows. Two words, never a sentence; never a word in the dictionary; never a name; admins see the list under Style guide and can remove any pair, and a guest sees none of it. Pairs that at least three workspaces independently accept go to a review queue the engineers read, where a human writes the rule and a blind test decides; nothing is applied automatically. Fourth, how you write, as a handful of numbers: sentence length, how often you use contractions, how you sign off, which English you spell in, and a short list of words you use often. It is kept in your browser only, listed in full under My settings so you can see every number and strike any out, and sent nowhere on its own; the only time any of it leaves is inside a rewrite you press the button for with the Rewrite in my voice switch on, as up to four short phrases you can read before you press. Quill does not build a profile of you as a person, keeps nothing you wrote, and no model is trained on anything you write.
• Your writing voice, and what it is made of. Quill notices how you write: sentence length, contractions, sign-offs, spelling. What it keeps is a handful of numbers about that, in your own browser, per workspace: average sentence length and how much it varies, how often you contract (do not against don’t), whether you say please, write as we or as I, use exclamation marks or the Oxford comma, spell the British or the American way, which sign-off you favour, and up to twenty words you use often. Those twenty words are the one place any words appear, listed on their own, never a sentence. The numbers are running averages, so there is no log of what you wrote to average from. It is sent nowhere on its own. Tick "Rewrite in my voice" on the Quill panel and a rewrite you ask for carries up to four short phrases made from the numbers, the most telling first, such as "prefers short sentences; uses contractions; signs off with Kind regards; British spelling", in a field the server treats as a preference and never as text to be rewritten; untick that and nothing is added. My settings on the Quill page shows every line of it, lets you untick any element so it is left out, and has a Forget my voice button that removes the whole thing. Nothing is trained on it, nothing about it reaches a server on its own, and it does not follow you to another computer. That is what the sentence above means in practice: this is not a profile of you, it is a description of your sentences, and it stays with you.
• Fix with Quill: the deeper AI proofread that corrects everything at once while respecting your dictionary and style guide, shown as a marked-up comparison before you apply it.
• Rewrite and tone: clearer, shorter, expanded, more formal, friendly, confident, direct or paraphrased; tone reads tell you how the message lands before you send it.
• Compose and reply: describe what you need, pick a language or keep the workspace drafting language, and get a ready-to-send draft (rewrites, tone and summaries always follow the language of your text). Drafts never invent numbers or names; anything unknown arrives as a [placeholder].
• Translation: thirty-plus languages in the picker, or type any language name.
• Snippets: the phrases your team types every day, saved once. Type the shortcut with a slash in front of it, press space or Tab, and it expands into the full text: /thanks becomes your thank-you paragraph. It works in every text box, every input and inside Drive documents, not only in the assistant panel. Shared snippets are managed by admins and available to everyone; personal ones are yours alone. An unknown shortcut is left exactly as you typed it.
• Style guide (admins): tone, audience, say-this-not-that terms, banned words and an optional replacement for each one, so a banned word becomes a one-tap fix instead of only an underline, the kinds of suggestion the workspace wants, the workspace drafting language for compose and replies, a reading-level target the live meter tracks, the watch-by-default switch and the daily AI allowance per person.
• Usage (admins): how much the team uses Quill, which actions, and how many suggestions are kept; and, since v1.288.0, suggestions by outcome: for each kind of suggestion and each rule, how many were accepted, ignored or added to the dictionary. The percentage is accepted out of accepted plus ignored. A suggestion nobody acted on is in neither number, because nothing about it ever left the browser, so read it as "of the ones people decided about".

Every AI action is counted and capped, and you can see the count. The buttons you press (Fix, Rewrite, Tone, Translate, Compose, Reply, Summarize) draw on a per-person allowance: a short burst limit, an hourly one, and a daily cap the admin sets on the Style guide tab. Whole-sentence checks as you type draw on their own hourly allowance per person and on nothing else. They used to share the buttons' allowance, and because they run without anybody pressing a button they drained it in about the first half hour of typing, after which every Fix and Rewrite was refused for the rest of the hour with nothing on screen to say why; that was corrected on 6 September 2026. The shared AI budget of the platform sits above both. The Studio tab on the Quill page and the assistant panel show what you have used of each ("You have used 14 of 200 AI actions today", "Sentence checking is on. It has sent 61 sentence checks this hour and pauses at 400"), read from the server rather than guessed. When a limit is reached Quill says which one, when it frees up, and that your text is never touched. CORRECTED: until 1 September 2026 sentence checks were recorded nowhere at all. The usage table refused the new kind of record, so the analytics on this page and anything a workspace would be billed against could not see them, while the sentence checks themselves went on answering. The recording is deliberately best-effort, so that a bookkeeping failure can never cost somebody their answer, which is why it went unnoticed for twelve releases. The underlines and the readability meter keep working whichever limit is reached, because they run entirely on your device.