38: Every-Day Writing Style - transcript
Liam: The QUILL Cast, episode thirty-eight. I'm Liam. Today we teach Quill to write in your voice: every-day writing style, the personal style guide that conditions the AI without ever sending your words anywhere you didn't choose.
Jessica: I'm Jessica. Episode thirty-seven, the AI toolkit, was the workbench: built-in tools, custom instructions per document, the Ask Quill shortcuts, and the standing instructions that travel with every AI verb. Today is the next bench over: a different lever, one that works on rhythm and tone rather than rules.
Liam: Right. Custom instructions are things you tell Quill, in plain prose, I write UK English, I prefer short sentences, never the word utilize. They're a list of do's and don'ts, and they apply everywhere AI generates text for you. The style profile is different. It's an inductive mirror: you show Quill samples of how you write, and Quill distills them into a guide that quietly conditions the assistant to sound like you.
Jessica: And the most important sentence in this whole episode: nothing about how the style guide works requires your words to leave this machine. The samples live in your local app data. The guide is built by the on-device model you already met in episode thirty-four. When the assistant generates text in your style, the assistant's the only thing doing the generation, with the style guide prepended to its prompt, and the guide is local.
Liam: The architectural name for this is prompt-based style conditioning. That phrase is in the docstring of the module that does the work, and it's worth pausing on, because it tells you exactly what's happening and exactly what's not.
Liam: What it is: the assistant receives a small block of text, your distilled style guide, ahead of your request, and is asked to match your voice. That's the entire mechanism. It is not training a custom model on your samples, and it is not a system that learns over time. It is one prompt, refreshed when you refresh it.
Jessica: What it isn't, and this matters for honesty: a future option, mentioned in the same docstring, is training a Foundation Models LoRA adapter from the samples, and that is genuinely future. It is not built yet. If you hear someone say Quill fine-tunes a model on your writing, that is not true as of this recording. Today: prompt-based conditioning, on-device build, local storage, fully reversible.
Liam: And the safety story from the toolkit episode is the same here. AI never changes your document silently. Every style-conditioned generation still arrives as a preview you accept or reject. The style just shapes the prose, the proposal mechanism is unchanged.
Jessica: Where the file lives, because the listener asked last week: the style profile is one JSON file under your app data, in the ai subfolder, named writing-style.json. The shape is small: a boolean for whether the style is enabled, a list of samples, and the guide text itself. Atomic write, schema stable, no migration worries.
Liam: Verified against code: the path is built by a function called style_path in quill/core/ai/style.py, and it returns app data dir slash ai slash writing-style.json. So if you ever want to back up your style, that's the one file. If you ever want to move it to a new machine, copy that one file, drop it in the same place on the next install, you're done.
Jessica: The other small but important number: the module keeps at most twenty samples, and each sample is truncated to four thousand characters. So if you paste in a long essay, only the first four thousand characters are stored. The model sees the concatenation of all your samples, up to a total of eight thousand characters, when it builds the guide. Those caps aren't guesses, they're in the constants at the top of the module.
Liam: Which means: the more representative your samples, the better. Don't paste a single twenty-thousand-character opus, paste four or five short pieces that span the registers you actually write in. A short professional email, a couple of paragraphs from a longer piece, a message to a friend, even a blog post. Variety in the samples beats volume.
Liam: Now the part that pays for the whole episode: the actual walk-through. From the menu, AI, Train Writing Style. The dialog is titled Train Writing Style and it's a single accessible window with three logical regions.
Jessica: Top region: a checkbox that says Write in my style when Quill generates text. That's the on-off switch. When it's on, the style guide is applied to every generation. When it's off, the samples and guide are still saved, they just don't condition the output. So you can teach Quill your voice on Monday and toggle the conditioning off on Tuesday for a job that needs a different register, no rebuild required.
Liam: Middle region: a text box labeled Add a writing sample, and two buttons beside it. Add Sample, and Learn from Current Document. Add Sample takes whatever is in the text box, trims whitespace, stores it as a sample, and clears the field. Learn from Current Document grabs the entire current document and adds it as a sample in one motion, very useful if you open an old piece of your own and want to feed it in.
Jessica: Below that, a small label, Samples collected, with a count. The module has a small nicety: every time you add a sample, either by adding or by learning from the document, the dialog automatically starts a build, so your style guide keeps itself fresh. You don't have to remember to press a build button after each add. The build runs on a background thread, the dialog announces Analyzing your writing, and when it finishes it says Style guide ready and enabled, and the on-off switch flips on for you automatically.
Liam: The build button, Build / Update My Style, is the manual override. It's also the smart default: if you press it with no samples in storage and the text box empty, the dialog falls back to learning from the current document. So Build just works. The docstring of the file is explicit about it: if nothing has been added yet, fall back to the sample field, then the current document, so Build just works.
Jessica: The bottom region: a read-only text box labeled My style guide, which shows you the current guide. So you can always see what the model distilled. If the guide is empty, the on-off switch is meaningful but the actual style work isn't happening, only enablement is. If the guide has text, that's the conditioning that will be prepended to prompts.
Liam: And a small accessibility detail we care about: the sample field is named Writing sample in the accessibility tree, the guide field is named Style guide, and the dialog is centered on the parent and uses the standard modal contract. The build runs on a worker thread, the results come back through CallAfter, the dialog never blocks the UI, and the screen reader gets discrete announcements at each step: Sample added, Analyzing your writing, Style guide ready and enabled.
Liam: Now, the do-this-now beat, before we go further. If you have AI set up, take a moment right now. Open the AI menu, choose Train Writing Style. The dialog opens. Click Learn from Current Document on whatever document you have open. If you don't have a document open that represents your writing, paste two or three short paragraphs of yours into the Add a writing sample box and click Add Sample, twice or three times. Then press Build / Update My Style. Pause the audio. The dialog will run, announce progress, and present a guide. Come back when you can see the guide text. We'll be right here.
Jessica: And the part to look for when you come back: the guide is supposed to be five to eight short bullet points. The prompt the module uses literally says produce a concise STYLE GUIDE, five to eight short bullet points, capturing their voice: tone, formality, typical sentence length, vocabulary, punctuation habits, and any quirks. Return only the bullet points, no preamble. That's the contract the on-device model is given. If your guide came back as a long essay, that's drift, and worth reporting.
Liam: One honest correction while we're here. An earlier draft of this episode said the guide was up to ten bullets, because the on-device model is generous. The actual prompt in code says five to eight, period. Verified. Five to eight is what you should expect and what you should be satisfied with. If the model returns more, the module stores what comes back, the count is not enforced, but the model is being asked for the shorter shape.
Jessica: A quick word on what good samples look like, because the listener who emailed asked exactly this. Good samples are yours, not someone else's, and they are from the kind of writing you actually want Quill to help with. If you want Quill to help you write professional updates at work, feed it two or three of your past professional updates. If you want Quill to help you draft a novel, feed it a chapter you already wrote and like. If you want help with short, conversational newsletter prose, feed it three recent newsletters. The match between sample register and target register is the single biggest determinant of guide quality.
Liam: And a quick word on what bad samples look like, because it's the inverse. Bad samples are inconsistent in voice, like pasting a tweet, a legal contract, and a bedtime story. The guide that comes out averages all three, and averages of inconsistent voices are bland. Bad samples are also overlong single pieces, where the model only sees one part of your range. And bad samples are heavily edited, like text that has been run through another AI, because the model will then be imitating the previous AI, not you.
Jessica: If you have an old document of yours that is mixed, with some sections that sound like you and some that don't, you have two clean options. First, just paste the whole thing and let the model find the dominant voice, which works if the dominant voice really is you. Second, take two minutes to copy out only the passages that most sound like you, paste those as separate samples, and discard the rest. The second approach is more work but more honest, and we recommend it for the first time you set this up.
Liam: When the guide is built, two more things happen automatically, and you should know they happened. The on-off switch at the top of the dialog turns on for you. And, back in the main window, the application re-applies the style to the assistant immediately, so the next thing you generate, whether that's a quick rewrite from the AI menu or a turn in Ask Quill, already sees the new guide.
Jessica: The wiring for that is in main frame, in the function that opens the dialog. After the dialog closes, the application calls a small helper that reads the just-saved profile and pushes the preamble into the assistant. The assistant stores the preamble, and from that point on, the preamble is prepended to every generation. It's the same single sentence plumbing we showed in the toolkit episode for custom instructions: instructions first, style second, then your request.
Liam: A subtle but important piece of that preamble, and the part that ties style to safety. The style preamble text is not a license to shorten or skip. The module is explicit, the preamble says, do not use it as a reason to shorten, omit, or skip content. Always fully complete the requested task at the length it deserves. So your style guide changes how Quill writes, not what Quill writes. A short style does not become a short answer that loses the point.
Jessica: Which means: the style profile is not a substitute for clear instructions. It's an amplifier of your instructions. If you ask for a two-hundred-word status update, the style guide shapes the prose, the length comes from your instructions. If you give no length hint, the assistant does the right length, and the style guide shapes how it reads.
Liam: What the style is allowed to influence, straight from the preamble: tone, word choice, and rhythm. Three things. Not argument structure, not facts, not accuracy. If a generated answer is wrong, the style profile didn't make it wrong, the underlying answer was wrong. If a generated answer is too long, the style profile didn't bloat it, you asked for too much or the model overran. The style profile is the voice; the rest is the content.
Jessica: And the same correction in the other direction: if the assistant is too formal for the voice you wanted, the style guide is one likely cause. Open the dialog, read the guide, see if it captured formality higher than your samples show, press Build / Update My Style again with a couple of more conversational samples added. The guide is regenerated from the current samples, not appended to. You have a refresh button, and it works.
Liam: The full edit-clear-save cycle, for completeness, because the listener who asked about it last week wanted to be sure. Clear Style empties the samples, empties the guide, flips the on switch off, and saves the empty profile immediately, even before you press the dialog's Save button. The dialog announces Writing style removed. So you can fully reset, locally, in one click, and the next generation runs without any style conditioning at all.
Jessica: And the JSON file on disk after a clear is the smallest possible: enabled false, samples empty list, guide empty string. There is no zombie state. If you want to be paranoid, you can also delete the file directly, the loader returns a default empty profile if the file is missing or unreadable, and the next save recreates it.
Jessica: Now the comparison that matters: style profile versus custom instructions, because they look like they overlap and they actually don't. Custom instructions are declarative, you write rules, the assistant follows them. Style profile is inductive, you write freely, the model extracts rules, the assistant follows the extracted rules. They compose.
Liam: The way they compose, ordered: instructions first, style second, then your request. That ordering is the order the assistant sees in its prompt. So if your instructions say never use the word utilize, that rule lands before the style guide is even read, and the style guide is asked to honor it. If your style guide says you use a conversational tone, that lands next, and the assistant tries to combine: no utilize, conversational tone, then the actual request.
Jessica: The practical takeaway. If you have a hard rule, like a spelling, a banned term, a required terminology, put it in custom instructions. If you have a feel, like the rhythm, the warmth, the formality, the punctuation habit, put it in the style profile. If you have both for the same thing, instructions win, that's the design.
Liam: And one more thing worth knowing: the custom instructions and the style profile are independent toggles. You can have custom instructions on, style profile on, both off, or any combination. The Apply Style to Assistant call in main frame sets both, so flipping the AI master switch off in the AI Hub will also stop the style from being applied, which is the right safety default.
Liam: A note on the relationship between this episode and the AI toolkit episode, episode thirty-seven, which is where the listener just came from. The toolkit episode established custom instructions and the AI shortcuts, and it set the per-feature safety contract. Today is the second of the two AI personalization levers. Tomorrow is the third major lever, agents, plans you review before they run, which is episode thirty-nine.
Jessica: The series arc, where we are in it. Episode thirty-four was set up. Thirty-five was Ask Quill. Thirty-six was the AI library and skills. Thirty-seven was the toolkit. Today, thirty-eight, is style. Tomorrow, thirty-nine, is agents. After that, episode forty opens part seven, the accessible vault, and the rest of the series, from forty through fifty-four, covers vault, story studio, glow, braille, quillins, trust, and the two power-user episodes that close the course.
Liam: Fifteen episodes to go after today, fifty-four total in the series. We named that number in the opening of episode one, and we still owe you every one of them.
Jessica: Limits of prompt-based style conditioning, spoken honestly, because honesty is the house rule. The on-device model that builds the guide is not a frontier datacenter model. It's the same model that powers your everyday AI actions. The guide it produces is good, often surprisingly good, but it's a distillation, not a measurement. Two authors with very similar samples can get slightly different guides, and a guide that perfectly describes your voice on Monday may be subtly off on Tuesday after you add a new sample from a different register.
Jessica: What that means practically. Don't over-tune. If you have a working guide that sounds like you, leave it alone. Add new samples only when your writing has genuinely changed, a new job, a new audience, a new format. The build is cheap, but each rebuild is a fresh distillation, and frequent small changes produce guides that drift toward the average of your samples rather than the strongest signal in them.
Liam: A diagnostic for drift. Generate a small piece in your style, read it back. If it sounds like you on a tired day, the guide is fine. If it sounds like a stranger who is trying to sound like you, the guide has drifted. Open it, read the bullets, see which one is off, adjust your samples, rebuild. Two or three iterations is usually the most you'll ever need.
Jessica: And a final honest line. The style profile conditions Quill's outputs to match your voice, but it does not make Quill you. You are still the writer. Quill is still the assistant. The profile just narrows the gap between the prose Quill offers and the prose you'd write yourself, so the accept-reject step is faster and the rejections are smaller. That's a real win. It isn't magic. It's a good tool, used well.
Liam: Three quick mistakes we have watched listeners make, so you don't have to make them. Mistake one is treating the style profile as a content filter. It is not. It is a voice filter. If you ask for a one-page summary and the style guide says you use long sentences, you will get a one-page summary written in long sentences, not a paragraph. Length comes from the request, voice comes from the profile.
Jessica: Mistake two is enabling the style on every kind of writing you do, then wondering why your technical documentation and your warm newsletter are starting to sound the same. Different registers want different voices. If you write in two distinct registers, professional and creative, build the profile for the one you write most, and toggle the switch off when you are doing the other. You can always keep a second profile in mind, hand-built in your head, by reading the guide, then turning the switch off when you want the unstyled voice.
Liam: Mistake three is rebuilding the guide every time you tweak a sample. The model doesn't average samples over multiple builds, it looks at the current set of samples and produces a guide. So a new build replaces the old guide entirely, it doesn't merge. If you keep the old sample that worked, add one new one, and rebuild, you may get a guide that has lost a bullet you liked, even if you did not remove the sample. That's not a bug, it's how distillation works. The fix is to be deliberate. Decide what you want the guide to say, select the samples that produce that, and build once.
Jessica: And the underlying skill, which is bigger than style profiles. Learning to direct an AI by adjusting its inputs is a craft. Style is one input. Custom instructions are another. The choice of model is another. The choice of action is another. The skill of stacking these inputs to get the output you want, that is what the rest of this part of the course is quietly building toward.
Liam: Homework, four steps, in order. One: open the AI menu, choose Train Writing Style, click Learn from Current Document on a piece of your own writing, watch the dialog build the guide. Read the guide. Does it sound like a fair description of how you write? If yes, leave it. If no, add one more sample, rebuild, compare. Two: toggle the on-off switch off, generate a short paragraph with Ask Quill, listen to the voice, then turn it back on, generate again on the same request, listen to the difference. Hear the conditioning, don't just trust it.
Jessica: Three: write a single sentence in custom instructions, something specific you actually want honored, I write UK English, I never use the word leverage as a verb, whatever is true, save it, generate a paragraph with both custom instructions and style profile on, confirm the instruction is honored and the style is present. Four: when you have a guide you trust, make a one-line note somewhere safe with the path to the writing-style.json file, so if you ever reinstall or move machines, you know exactly which one file to copy.
Liam: Next episode, thirty-nine: agents, plans you review before they run, the accessibility tune-up that makes the whole idea usable, and the moment this part of the course stops being about features and starts being about delegation.
Jessica: I'm Jessica.
Liam: I'm Liam. Teach Quill your voice, then trust your own.
Jessica: The QUILL Cast is a fifty-four-episode audio course on QUILL. This was episode thirty-eight. Until episode thirty-nine, write well.
Liam: And read the guide out loud, once. It tells you more than you think.