35: Ask Quill and Everyday AI - transcript
Jessica: The QUILL Cast, episode thirty-five. I'm Jessica. Setup is done, we are through the wizard, your engine is configured, your key is in, and your writing has not changed yet. Today is the day it starts to. We open the AI menu, we meet Ask Quill, we run the everyday writing actions, and we build the habit of reviewing every suggestion before it touches a paragraph.
Liam: I'm Liam. Quick recap of episode thirty-four, because we want the bridge to be clean. Last time Jessica walked you through the AI setup wizard. Three paths: connect a cloud provider, plug in an AI agent you already pay for, or pick Not Right Now, which is a fully supported answer. We named the three rules that govern every AI feature in QUILL. AI is off until you turn it on. You choose who, if anyone, sees your words. And the safety story in one sentence: AI never changes your document silently, every edit arrives as a preview you accept or reject, and an accepted change is one undo step. Those rules are not aspirational. They are enforced in the code, in the menu gates, and in the dialog contract. We promised that today's episode would turn that configuration into habits. That promise starts now.
Jessica: And one honest correction before we go any further, because honesty is the house rule. The earlier short version of this episode, the one that lived in this slot before today's rewrite, named Ask Quill as the centerpiece, which is right, and it named a list of quick actions, Rewrite, Summarize, Expand, Continue Writing, Fix Grammar, Generate Table of Contents, Translate, and Document Q and A, and that list is also right. What the earlier version did not surface was the deeper architecture underneath. Ask Quill, the chat, and every quick action, Rewrite, Spell Check, Grammar, Thesaurus, Translate, Document Q and A, all share one engine. That engine is in quill core ai chat session, and the unification is recent. There used to be multiple chat surfaces; they have been collapsed into one conversation that knows your document. Today's episode explains what that means for how you actually work.
Liam: The core idea to hold in your head, before we touch a menu, is the layering. QUILL has a long history with local engines, spell check, basic grammar, the classic thesaurus, going back to early episodes in this series. AI Spell Check, AI Grammar and Style, and the AI Thesaurus are not replacements. They are a correction tier above the local engines. Local catches what rules can see. AI catches what only a model can see, real-word errors like their versus there in context, agreement across long sentences, a synonym that fits the meaning rather than the spelling. The local engines stay free, instant, and offline. The AI tier is opt-in, runs only when you ask, and is gated by the same review contract as everything else. The layering is intentional, and it is the reason QUILL works fully with AI off, the same way it always did.
Jessica: On top of that engine, QUILL exposes two surfaces. The conversation surface, Ask Quill, opens a real chat, multi-turn, persistent within a session, knows your selection and your document as context, and is the place to interrogate, refine, and explore. The quick-action surface, the menu items under Proofread and Transform Selection and Translate, is for when you do not need a chat, you just need a verb. Tighten this paragraph. Summarize this chapter. Continue from where I stopped. Fix grammar. The two surfaces share one brain. That is the architectural truth. From your seat it looks like a menu with chat at the top and verbs below. From the code it is one engine, many faces, all subject to the same preview rule.
Liam: Let's open the door, and the door is in the AI menu. In the build I am reading from, the AI menu opens with Set Up AI if you have not run the wizard yet, then Ask Quill and Ask Quill by Voice, then a divider, then the everyday actions grouped under submenus like Proofread, Transform Selection, Translate, Read Aloud, Transcribe Audio, and More. The Library and the Hub sit further down. Ask Quill and Ask Quill by Voice are at the top because they are the front door. Every other AI thing in the product is reachable from chat, from a quick action, or from the Library, and that ordering is deliberate. Conversation first, verbs second, building blocks third. Today we walk all three.
Jessica: Here is your do this now beat, and we want you to actually do it, with the audio paused, not just nod along. If QUILL is open, open a document you do not mind experimenting in, a scratch file, a notes page, anything with at least a paragraph of your own prose. If QUILL is not open, that is fine, you can do this after the episode, the rest will still make sense. Place your cursor in the middle of a paragraph, with no selection, just the cursor sitting inside the text. From the AI menu choose Ask Quill. The dialog opens. Before you type anything, take ten seconds to listen to the screen reader, or to look at the dialog structure. Notice the input field. Notice the model label, which is the engine you set up last episode. Notice that the conversation is empty, fresh, ready. Now type exactly this: what is unclear in this paragraph, and press enter. Wait for the answer. Read it. Decide for yourself whether the answer is useful. Accept, reject, or ask a follow-up. There is no wrong answer. Press play when you are back.
Liam: That little exchange, question, answer, follow-up, is the whole point. The model is not the product. The conversation is the product. The model is just one participant in it, a fast one, a cheap one, sometimes a wrong one, and your job is to be the editor in the loop. QUILL's code expresses this directly. The chat session module, chat session dot py, models the conversation as a list of turns, with three roles, system, user, assistant, and an auto-compaction step that summarizes the older head of the conversation when it grows past the budget. What this means for you is that long Ask Quill sessions do not silently forget the early turns. They compress them, the model is told the gist, and the recent turns stay verbatim. You can run a real conversation, ten, twenty, thirty turns, and the context does not fall off a cliff.
Jessica: A second thing you may not have noticed yet. Ask Quill passes your document and your selection to the model as context, but the code is defensive about this. The assistant class in quill core ai assistant dot py has a clamp message function that bounds a single user message so a huge paste cannot exceed the model's context window and hang inference. It also tries shrinking document context budgets, six thousand chars, three thousand, twelve hundred, zero, until the prompt fits. If a model is smaller, Ask Quill adapts instead of failing. This is a quiet kindness, and it is why a free local model still gives you a working conversation instead of a hung dialog. The robustness is built in, not advertised.
Liam: The quick actions, the verbs, are where most people will spend most of their AI time, so let's walk them honestly. Rewrite, on the menu, Transform Selection, Rewrite Selection. Select text, run it, get a preview, accept or reject. The preview is delivered as a structured response, the model returns the rewritten text, no preamble, no "here is a rewrite:" because the prompt explicitly forbids that, and the dialog shows the proposed text with an accept and a reject control. That is the loop, and you can run it as many times as you like. Each run is independent. The model does not remember your previous rewrite. If you want a second draft, you ask Ask Quill, "give me a second rewrite of that paragraph," and the chat carries the context.
Jessica: Summarize. The prompt in the code asks the model to lead with the central claim, include the key supporting points, drop examples unless one is the clearest statement, target roughly one fifth of the original length, and return only the summary, no opener, no closer. Notice the design choice. The default ratio is one fifth, and the model is told that shorter is fine if the content is thin, longer is fine for dense technical material. That is a constraint that improves the quality of the output, because a model asked to summarize without a target length produces a wall of text. One fifth is a heuristic that works across register, formal, casual, technical, literary. You can still ask for a different length in Ask Quill, "summarize that in three sentences," and the chat will respect it.
Liam: Expand is the mirror of Summarize. Compressed notes to prose, the model is told to add concrete detail, specific examples, transitions, and to match the existing tone and register with precision. Importantly, the Expand prompt explicitly says, do not state facts you cannot verify, and to use qualified language, "this suggests," "one possibility is," when moving into inference. That is a guardrail against hallucination, and it is one of the reasons QUILL's expand is honest, the model is told that an expansion that invents facts has failed. Continue Writing is different again. It takes the document up to your cursor, or your selection, and asks the model to keep going. The model is told to use the existing voice, not break the fourth wall, not stop and ask a question. Useful for the moment you are staring at a blank half of a paragraph and need momentum.
Jessica: Fix Grammar is the everyday writer's best friend, and the one I would teach first. It runs AI Grammar and Style, the code is in quill core ai grammar check dot py, and the response is a list of structured issues, each with a category from grammar, punctuation, clarity, style, word choice, a suggested fix, a one-sentence explanation, and a forty-character context window. Each one is reviewable individually. You do not accept the whole list as a block. You see them one at a time, with the explanation, and you choose. That is the design, and it is the difference between a tool that improves your writing and a tool that quietly flattens it. The default custom instruction for grammar, which is in custom instructions dot py, is short and worth quoting in spirit: prioritize issues that genuinely impair comprehension first, then punctuation, then style. Preserve the author's voice. Never flag something as wrong just because you would phrase it differently.
Liam: The corrected tier. AI Spell Check, AI Grammar and Style, AI Thesaurus. The local engines from earlier episodes stay on by default and stay free, and that matters for a reason we should name. On a fresh QUILL, with no AI configured, F seven still spell-checks using a dictionary. The document workflow works fully with AI off. That is not a fallback, that is the architecture. When you turn AI on, you add a layer. The local engine runs first, fast, free, offline. The AI tier runs second, on what rules can't see. Real-word errors, their there they're, its it's, affect effect, caught in context. Agreement and clarity across long sentences. Synonyms ranked by meaning in the exact sentence, not by alphabetical distance. Spell Check returns a list of corrections, each with original, correction, and a forty-character context window so you can see what surrounds the word. The thesaurus returns six to ten synonyms, ordered from most to least similar, with a one-sentence note on how meaning shifts.
Jessica: A subtle thing about the thesaurus that the earlier version of this episode did not mention. The thesaurus is sentence-aware. You can pass it a context sentence, the sentence the word appears in, and the model is told to read that context carefully and suggest words that fit that exact meaning, not the word's other senses. That is why a QUILL thesaurus suggestion is usually better than a generic synonym list. The default custom instruction in custom instructions dot py is also explicit on this: rank from most to least interchangeable in this specific context. Flag meaning shifts that matter. Add a register note when it is useful. Six to eight well-chosen synonyms beat a list of twenty. Quality over quantity. That prompt is the reason the thesaurus does not flood you with options, and the reason the option it leads with is usually the right one.
Liam: Document Q and A is the long-form companion feature. Open it from the AI menu, point it at a long document, a sixty-page PDF you imported in episode twenty-two, a research report, a transcript from episode thirty-two, and ask a question. What are the deadlines mentioned in this report. Summarize the third chapter. Who is the decision maker on the budget line. The system prompt restricts the answer to the document. If the document does not address the question, the model is told to say exactly, the document does not address this, and to not speculate. That refusal is not a failure. It is a feature. The Document Q and A custom instruction makes the same point in different words: draw every answer from the document, do not use outside knowledge, quote short relevant passages, keep answers tight. This is the long-form reading assist that the early version of this episode promised, and it works on the free local model for short documents and on a cloud model for the long ones.
Jessica: And Translate. The earlier version said Translate, selection or document, and that is accurate, the menu items are Translate Selection and Translate Document. The translation prompt, in the code, is a full professional translator brief, prioritize natural idiomatic expression, preserve the author's tone and formality, localize idioms and culture-specific references when a literal rendering would confuse a native reader, preserve all formatting exactly including Markdown headings, bullet lists, bold and italic markup, HTML tags, and code spans, return only the translated text. The supported languages list, in translation dot py, is long, and the list grows as the model roster grows. The point for the writer is that the translation is not a word swap. It is a register-preserving rewrite, and the formatting is preserved as part of the contract. If you translate a Markdown list, you get a Markdown list back, not a paragraph.
Liam: Custom instructions, and this is the underused superpower. The custom instructions module, custom instructions dot py, lets you pin standing rules that every AI feature honors. Spell Check, Grammar, Rewrite, Summarize, Expand, Table of Contents, Translate, Thesaurus, Document Q and A, Research, and the chat itself, Ask Quill, all have default system prompts and an optional user override. You open the custom instructions editor, you set I write UK English, I prefer plain language, never use the word utilize, and from that point on, every AI feature in QUILL honors it. Spell check stops trying to Americanize your spelling. Grammar stops suggesting Oxford commas if you have turned them off. Rewrite stops inventing enthusiasm you did not ask for. Set once, honored everywhere. That is the difference between a tool and your tool.
Jessica: And there is a second, separate layer of standing rules, the writing instructions, and we want to name it because it is easy to confuse with custom instructions. Writing instructions live in plain Markdown files you can open and edit. The global file is at app data, ai, writing instructions dot md, and it applies to every document. A document sidecar, named the document name plus a quill instructions suffix, applies to that document only. The two are concatenated, global first, document second, so a document can refine the house style. Both are visible, user-owned, and live reloaded. Where custom instructions are per-task rules, writing instructions are the durable voice and house style. The architecture is in writing instructions dot py, and the rule is simple: instructions lead, then style, then the task. Both are visible, user-controlled, and on disk, not hidden in a settings blob.
Liam: This is also a good place to mention the Ask Quill companion sidebar that you may have noticed in the screenshot. It is the chat surface, anchored to the right of the document, that lets you keep your cursor in the document and your conversation in the sidebar. The chat session in chat session dot py is provider-neutral and serializes to disk. You can close the document, reopen it tomorrow, the sidebar restores. The transcript is there, and the model is the same one you set up in the wizard. None of this is magic, it is the chat session class being persistent and the same code path powering a dialog and a sidebar, and it is the right shape for everyday work.
Jessica: Real workflows, the kind that earn their place in a writing day. The morning email, draft rough, three lines, run Fix Grammar, accept two suggestions, send. Ninety seconds total. The report polish, open the draft, run Ask Quill with the prompt, what is unclear in this section, get an answer, disagree with one of the three suggestions, accept two. The reading assist, open a sixty-page imported document, open Document Q and A, ask, what are the deadlines mentioned, get an answer, jump to the source. The long-form draft, hit a wall mid-paragraph, select the last two sentences, run Continue Writing, read the model's continuation, accept the part that lands, reject the part that does not, move on. In every case, the human stayed in charge of meaning. The AI handled labor. That is the pattern, and once you see it, you see it everywhere.
Liam: A habit recommendation from us, sincerely, and this is the part of the episode that earns its keep. Reject things. Regularly. Not out of spite, out of calibration. The reject button is the dial that keeps your judgment sharp and the tool in its lane. An AI feature you have never said no to is a feature you have stopped supervising. The earlier version of this episode made the same point, but I want to be more concrete. In a typical week, a healthy rejection rate for AI suggestions is between twenty and forty percent. Not because the model is bad, because you have taste, and taste requires saying no. If your rejection rate is near zero, ask yourself whether you are paying attention. If it is near one hundred, ask whether you are using the right model for the task.
Jessica: And one more honest correction, because the previous version got a small thing wrong. The earlier script said that the AI pass replaces the local engines. It does not. The local engines are still on by default, still free, still instant, still offline. The AI pass is layered on top, gated by review. If your AI is off, or you have not run the wizard, or you are in safe mode, the local engines still do the job they have always done. The layering is the architecture, and it is the reason QUILL is honest about what AI is and is not. AI is a tier, not a swap.
Liam: Let's talk about the safety contract one more time, because today is when the contract goes from theory to habit. Every AI feature in QUILL, Ask Quill, the quick actions, the corrected tier, the sidebar, returns a structured response. The diff review module, diff review dot py, is the surface that turns model output into a preview you can accept or reject. The accepted change is one undo step. The rejected change is nothing happened. There is no silent write, no auto-accept, no "trust me" mode. The dialog contract, the apply modal ids function, makes sure the preview dialog has a clear accept, a clear reject, and a screen-reader announcement for the result. You are the last link in the chain, and the chain is built so that the last link is load-bearing. That is not a slogan. It is the code.
Jessica: The honest coda on what this episode does not cover. We have not talked about the AI Library, where you save prompts and build skills, and that is episode thirty-six, next time. We have not talked about agents with reviewable autonomy, that is the episode after. We have not talked about the everyday writing style, which is the part where QUILL learns your voice from your own samples. Today is chat and verbs. Tomorrow is workshop. The layering is the same, the review contract is the same, but the surface area grows, and the next two episodes are where the surface area earns its depth.
Liam: Homework, four steps, designed to be done this week, not in a single sitting. One: run Rewrite on a paragraph you wrote, listen fully to the preview, then reject it. The point of the reject is to feel that the no is real, the dialog closes, the document does not change, the undo history is unaffected. You now know the machinery is honest. Two: run it again on a different paragraph, accept the rewrite, undo it with control Z, and redo it. The point is to feel that accept is reversible in one keystroke. You now trust the contract, not just believe in it. Three: have one genuine Ask Quill conversation about a real draft, at least three turns, where you follow up on the model's answer. The point is to learn that the chat is the surface for interrogation, not just for one-shot questions. Four: set one custom instruction that is true about your writing. UK English, plain language, no em dashes, whatever. Save it. Notice on the next Spell Check that your dialect is honored. The custom instruction just became infrastructure.
Jessica: And one bonus step, free of charge. Open the AI menu, look at the bottom, and find Use Artificial Intelligence. That is the master switch. If you ever want the quietest possible QUILL, uncheck it. The menu grey out, the chat and the quick actions become inert, the local engines still work. The product you are holding has the off switch at the surface, not buried in a settings dialog. That is not an accident.
Liam: Next episode, the workshop. Episode thirty-six, the AI Library, where you save prompts, build skills with multiple steps, and climb the promotion ladder from a one-off chat command to a reusable skill you can share. The Library is the place where everyday AI becomes your AI, where the prompts that earned their keep last week are saved this week, where one prompt becomes three steps becomes a packaged skill. The chat and the verbs from today are the raw material. The Library is where that material becomes a craft.
Jessica: A quick number, for the show notes. The QUILL Cast is a fifty-four-episode audio course. This is episode thirty-five. We are in the home stretch of the artificial-intelligence part. The AI Library and the AI Toolkit are next, then every-day writing style, agents with reviewable autonomy, accessible vault basics, vault power, the Story Studio work, the GLOW audit work, the GLOW for files work, braille production, the Quillins and the console, and the trust and community finale. Twenty episodes after this one. The pace from here is one topic per episode, with deep dives where the surface area earns them, and today's deep dive on Ask Quill and the everyday writing actions is the foundation the rest of the AI part is built on.
Liam: I'm Liam.
Jessica: I'm Jessica. Propose, review, dispose, in that order. Until episode thirty-six.