37: The AI Toolkit - transcript
Jessica: The QUILL Cast, episode thirty-seven. I'm Jessica. Today the workbench, the AI toolkit, the set of always-there levers that turn the AI menu from a list of commands into a workshop.
Liam: I'm Liam. A small honest correction before we go further. Last episode was actually thirty-six, the AI Library, prompts, and skills, the build-along where a status-update prompt climbed the first two rungs. We promised agents next. Agents do come next, but the work in between, the always-there toolkit, earned its own episode, this one. So today is rung one and a half: the everyday tools, the safety story they share, and how to make the AI sound like you without ever training a model.
Jessica: The recap of thirty-six in one breath: AI Library is one tabbed home, under the AI menu, with Prompts, Skills, and Agents. Prompts are reusable templates with variables, single-step transforms. Skills are multi-step workflows, each step feeding the next, like our status-update example, summarize, restyle, prepend subject. Agents are the third rung, and we parked them.
Liam: That ladder is still standing. Today we do not climb it. Today we walk the floor of the workshop and meet the tools that have been quietly doing their job underneath every rung. After today, the next two episodes are rung three, agents with readable plans, and then the personal style mirror, every-day writing style, which is episode thirty-eight.
Liam: The toolkit has four families. One, the built-in catalog of pre-written tools you can run on any selection: rewrite, summarize, expand, fix grammar, continue writing, generate a table of contents. Two, the always-on safety story they share. Three, the standing instructions that travel with every AI verb, the global and per-document custom instructions. Four, the Ask Quill shortcuts, the one-off chat for when a tool is not quite the right shape. We'll walk each.
Jessica: The first family lives in the AI menu under a submenu called Transform Selection. There are six entries. Rewrite Selection, Summarize Selection, Expand Selection, Continue Writing, Fix Grammar, and Generate Table of Contents. Each one is a real, named, keyboard-reachable verb. They are not the Library, they do not get saved, they are just there when you need them, and they do exactly what their labels say. If you've ever used word processors, you have a mental model: a button, a result, a preview.
Liam: Right. And notice what they all share. None of them edit the document silently. You pick one, the AI reads the current paragraph, selection, or document, drafts a revision, and a preview window appears. The preview has three buttons by default: accept, reject, and the always-there undo, which is also control Z, your normal QUILL undo, immediately after accept. That is the entire safety contract. Propose, preview, decide. We hammered it in episode thirty-four, the setup episode, and it has not changed.
Jessica: Two of these verbs deserve a moment because they fix problems a lot of people give up on. Continue Writing reads the cursor context, paragraph or document, and proposes what to write next. It is the unsticker, the answer to staring at a blinking cursor. Fix Grammar is the same job, but specifically targeted at grammar, so it is faster and cheaper. If you find yourself writing the same kind of cleanup week after week, both of these are worth a keybinding.
Liam: The third is less obvious. Generate Table of Contents runs over the whole document, looks for headings, and proposes a Markdown list with anchored links. It is the kind of feature that lives behind paywalls in other apps, and here it is a single-shot transform. Run it on a fifty-page manuscript, the preview is a TOC, accept it, you have a TOC. There is no learning curve. The verb knows what the shape is.
Jessica: Second family, the safety story. We keep saying it because it is the one rule that never bends, and it is the difference between AI you can trust with your writing and AI you cannot. The rule has three parts. One, the AI never edits the document silently. Every change arrives as a preview you accept or reject. Two, an accepted change is one undo step. Control Z takes it back, exactly like any other edit. Three, you see what the AI saw. The preview window shows the source text, the revised text, and the model name. If something looks off, you know what to look at and where to push back.
Liam: There is a quiet extension of that rule that matters for screen-reader users. The preview window is fully accessible. It uses standard list and text controls, every button is keyboard reachable, and the announcement when the preview opens tells you the model and the count of changed lines. So the experience is not second-class, it is the same first-class safety story, just voiced instead of seen. This is the contract from episode twenty-six, the way the preview pathway works, and it applies to every single one of these tools.
Jessica: Third family, the standing instructions. This is where the toolkit gets personal, and it is the part I want you to pause for, because the next minute is the single highest-leverage thing you will do this episode.
Liam: There are two instructions surfaces. Global and per-document. The global one lives at AI menu, then More, then Writing Instructions. The first time you open it, QUILL creates a friendly template, three blank lines: audience, tone, words to avoid. You edit it like any other document, save it, and from that moment on, every AI verb in QUILL reads those rules and obeys them. No more pasting the same preamble into every prompt. No more teaching Quill your house style once per session.
Jessica: The per-document version is the same idea, but scoped. If a particular document is for a different audience, a technical brief versus a blog post, you can write a small sidecar instructions file that travels with that document. QUILL reads both, the global first, then the per-document overrides, so you can have sensible defaults everywhere and project-specific exceptions anywhere.
Liam: And the standing instructions travel with every AI verb. The Transform Selection verbs, the Ask Quill chat, the agents next week, the style mirror the week after. One file, every feature. That is the leverage.
Liam: Okay. Here is the do-this-now beat, and I mean it. Pause the audio, open the AI menu, choose More, then Writing Instructions. QUILL creates a template if you do not have one yet. Write three lines. Audience, for example, a manager who wants the headline first. Tone, for example, plain language, no jargon. Words to avoid, for example, never utilize, never leverage as a verb, never in conclusion. Save the document. Come back when you have three real lines. This is twenty minutes of value per week for the rest of the year.
Jessica: While you are doing that, a quick honest aside. We said in the show notes that you would customize the per-document instructions, too. That is in the standing instructions tab inside the AI Hub, the Instructions sub-tab, the one with the writing tasks list and the per-task override. The menu entry called Writing Instructions is the global file. Both matter, but the menu entry is the one to start with, because global rules travel with every document, and most of the time that is what you actually want.
Liam: Welcome back. The fourth family is Ask Quill, the one-off chat. AI menu, Ask Quill, or the Ask Quill by Voice submenu if you have a microphone and prefer to talk. Ask Quill is the shape to use when a built-in verb is too narrow, when you want a real back-and-forth, or when the question is about the document rather than a transformation of it. Summarize this chapter in three bullets for a meeting tomorrow, draft a follow-up email based on the last paragraph, what assumptions am I making in this argument.
Jessica: Ask Quill also has a memory across turns in the same session. You can ask a first question, see the answer, refine, redirect, and the model keeps the thread. That is not magic, it is just the chat shape, but the result feels different from the single-shot transforms. Use the transforms when you know exactly what you want done. Use Ask Quill when you want to think with the AI.
Liam: A note about voice. Ask Quill by Voice opens the same chat dialog, but with a spoken-input field and a text-to-speech output that uses whatever voice catalog you have set up, SAPI, Kokoro, Piper, eSpeak, the lot we covered back in episode twenty-one. If voice is not set up, the dialog still works, it just falls back to typing. The accessibility story is the same in both modes, focus and screen readers are first-class.
Jessica: And here is the connective tissue between the four families, the part that makes the toolkit feel like a toolkit rather than a list. Every one of these tools reads the same standing instructions. Every one of these tools obeys the same preview-then-accept safety story. Every one of these tools can be invoked from the AI menu, the command palette, or a keybinding. So once you have learned one, you have learned the shape. Pick a verb, set a context, get a preview, decide. The rest is detail.
Liam: Time for the code-verified walkthrough, since you should never trust a podcast that names menu items by feel. We are reading the actual menu builder in quill slash ui slash main underscore frame underscore menu dot py, lines around twenty eighty, where the AI menu is built. The first entry is always Set Up AI, with a special start-here label until AI is configured, then a separator, then the conversation front door, Ask Quill and Ask Quill by Voice, then a separator. After that, in Advanced mode, we get the Accessibility Tune-Up, the Concierge, the Rewrite and Improve submenu, and the Run Agent submenu. The transforms live further down under Transform Selection. The proofread verbs, grammar, spell check, live under Proofread. The Read Aloud verbs live under Read Aloud. The transcription verbs live under Transcribe Audio. The Document Q and A, AI Thesaurus, Train Writing Style, and Writing Instructions live under More. The AI Library is its own top-level entry, then the AI Hub, then a separator, then the Use Artificial Intelligence master switch, and finally the Show advanced AI features toggle.
Jessica: So the menu is bigger than it looks, and that is on purpose. The four-pillar design, which we covered way back in the AI menu redesign, is: conversation at the top, context actions next, transform and proofread in the middle, and management at the bottom. The complexity is bounded, though, because every entry follows the same preview-then-accept story.
Liam: And the agents folder, which is what we are building toward for next week, is real. We verified it against quill slash core slash ai slash agents. There are sixteen Markdown files, one per agent, each with YAML front matter, an id, a display name, a description, a risk level, a default scope, recommended file types, and a permissions block. The Markdown body is the agent's instructions, the system prompt. That is the authoring standard. The build's standards linter, python dash m quill dot tools dot agent underscore lint, enforces the format on every check-in. So when we say the agents are reviewable, we mean it literally. Open the file, read the prose, that is what the AI is told.
Jessica: Sixteen agents shipped today. Summarizer, Reviewer, Plain-Language Rewriter, Research Assistant, Math Tutor, Writing Companion, Meeting Notes to Action Items, PRD Architect, Release Notes Builder, GitHub Maintainer, Markdown Publisher, Accessibility Editor, Citation and Link Fixer, Code Doctor, Data Cleaner, and QUILL Concierge. Each one has a job you can describe in one sentence. Each one runs through the same Safe Editor Tool Gateway that the transforms use. Each one can be edited by hand if you disagree with how it was written.
Jessica: Now, a small but important honesty moment. The concierge agent, the one whose display name is QUILL Concierge, is a different thing from the Concierge menu entry you see in the AI menu. The Concierge menu entry is a context-aware action picker that suggests which agents to run, given your current file type and selection. The agent called QUILL Concierge is one of the catalog agents you can run directly, and its job is to answer what can I do here with a list of suggested actions. The naming is a little confusing, and we own that. The menu entry is the front door. The agent is the back-end brain behind the menu entry. Both are useful, and they are not duplicates of each other.
Liam: Another honesty moment, about the Run Agent submenu. It is a real submenu, in the AI menu, in Advanced mode, populated automatically from the agent catalog. If you install a new agent, it appears there. If you edit the display name in the Markdown, the menu updates after a refresh. The build enforces the standard, so the menu cannot drift from the catalog. This is the kind of plumbing that sounds boring until you need to change one of your sixteen agents, and then you realize it is the entire reason a custom agent workflow actually works at all.
Jessica: And a note about a previous-episode claim. In episode twenty-six we said the Ask Quill chat is a fully modeless dialog that stays open while you work. That is still true. We also said it remembers the thread across turns, which is also still true. What we did not say loudly enough is that the chat has its own model selector. In the AI Hub, the Engines tab, you can pick a different engine for Ask Quill than for the transform verbs, if you want, say, a local model for the transforms and a stronger cloud model for the chat. Most people do not need that, but the lever exists.
Liam: A short tour of the supporting submenus before we tie it together, because the toolkit is bigger than the six transforms and the chat. Under Proofread you get four verbs. Check Grammar with AI is a one-shot pass, ideal for outbound mail. Grammar and Style Check is the same idea but stricter, it asks for tone and clarity, not just comma placement. Spell Check is the AI version of the spell check you already know, useful for medical or legal text where a missed term is a real problem. Spell Check Interactive walks the document issue by issue, asking for each one before changing anything, so you stay in full control.
Jessica: Under Translate, two verbs. Translate Selection runs on what you have highlighted. Translate Document runs on the whole file, in chunks, and produces a new document in a target language. The translation can use your configured AI provider, or a local LibreTranslate instance if you have one running. The safety contract is the same, you get a preview, you accept or reject, and the original is untouched.
Liam: Under Read Aloud, the AI voice verbs. Read Selection Aloud speaks the current selection with the active voice from the catalog. Read Document Aloud speaks the whole document. Stop AI Reading halts an in-progress read, important if you started a long document and changed your mind. Export Document as Audio writes the spoken version to an MP3 file you can keep, the audiobook workflow, basically, in one verb. We covered the audio studio in episode thirty-three, this menu is the everyday on-ramp to the same machinery.
Jessica: Under Transcribe Audio, two verbs. Transcribe Audio File is the one that turns an audio file into a written transcript, with optional speaker diarization and optional translation to English, all in the same dialog. Transcript Actions runs the Listening Companion actions on whatever is currently selected or the whole document, the same actions that fire after a fresh transcription. That is the connection to episode twenty-three, the Listening Companion work. The Transcript Actions surface, which is also in the toolbar in some modes, is the always-there way to run them outside a fresh transcription.
Liam: And under More, the deeper tools. Document Q and A is the multi-turn chat scoped to a single document, you ask questions, the model answers from the text. AI Thesaurus is the find-a-better-word verb, but with the same preview shape. Train Writing Style is the one we will dive into next episode. Writing Instructions is the standing-instructions file we already covered.
Jessica: So pulling the threads together, the toolkit is six families, one shared safety contract, and one shared set of standing instructions. The transforms, six of them, do the everyday shape-changing jobs. The proofread submenu, four of them, handles the editing checks. The translate and read aloud submenus cover language and audio. The transcription submenu handles the listening companion workflow. The More submenu holds the deeper tools, Document Q and A, AI Thesaurus, Train Writing Style, Writing Instructions. And the AI Library, AI Hub, and master switch are the management layer.
Jessica: And the through-line is the preview. Every action that changes your document goes through the same preview pathway. You read the proposal, you accept or reject, and an accepted change is one undo. That is the entire safety story, told once, enforced everywhere. It is why the toolkit is a toolkit and not a pile of features. The shape is the same. The work is different.
Liam: A note for screen-reader listeners, because this matters. Every preview window in QUILL is built with the same dialog contract, the one we covered in episode twenty-six, the one the dialog inventory gate enforces. That means focus, keyboard order, labels, and announcements are consistent across every AI verb. If you have learned how to drive one preview, you have learned them all. The concierge, the transforms, the proofread, the Library inserts, the agent previews, all the same shape, all the same announcements, all the same keyboard.
Jessica: A note on series arc. We are now seventeen episodes into the AI part. Episode thirty-four was setup. Thirty-five was Ask Quill. Thirty-six was the Library, prompts, and skills. Thirty-seven is the toolkit, today. Thirty-eight is the personal style mirror, every-day writing style. Thirty-nine is agents, readable plans, autonomy you can review. Then we leave the AI part and start part seven, the accessible vault, beginning at episode forty. The total arc is fifty-four episodes, fifteen more after today, and the toolkit is the foundation that everything from here on rests on.
Liam: Time for the homework, four steps, designed to be done in one sitting. Step one, if you skipped the do-this-now beat, go back and write your three standing-instruction lines, audience, tone, words to avoid, save the global instructions file, and run any AI verb to see the rules in action. Pick something with a real audience in mind, a work email, a report intro, and notice how the rewrite changes shape immediately.
Jessica: Step two, pick one of the six Transform Selection verbs, whichever solves a real problem you have, and bind it to a key. The command palette, control shift P, takes you to the binding editor. If Continue Writing solves your stuck-cursor problem, bind it. If Fix Grammar saves you ten minutes a week on outbound mail, bind that. The point is to convert the menu entry into a one-keystroke habit.
Liam: Step three, open the Ask Quill chat, ask one real question about a document you are working on, and follow up with a refinement. The point is to feel the back-and-forth shape, not to get a perfect answer. Notice how the model holds the thread. Notice how the standing instructions are quietly applied without you having to repeat them.
Jessica: Step four, open the AI Hub, the Engines tab, and read what engines are currently registered. If you have a free key from episode twenty-five, it is in there. If you are running Ollama locally, it is in there. If you have nothing configured, run Set Up AI and pick the path that fits your hardware and your privacy needs. The point of this step is to know what is actually powering your toolkit, not to assume it is the cloud when it is your laptop.
Liam: Next episode, thirty-eight, every-day writing style, the personal style mirror. The standing instructions we wrote today are rules, do this, never that. The style mirror is different. It is inductive. You give QUILL samples of how you write, and it distills them into a guide that quietly conditions the assistant to sound like you, the rhythm, the cadence, the sentence length, the kind of metaphor you reach for. Same safety contract, same preview pathway, different lever.
Jessica: And the week after that, thirty-nine, agents, the third rung of the Library ladder. Plan-first, review-the-plan, approve-or-trim, then watch the agent work. That is the autonomy you can read before it happens, and it is where the toolkit turns into a workshop you can trust with whole documents, not just sentences.
Liam: One last thing. We said earlier that the four families share the same preview contract. We will say it again next episode, and the episode after that, because it is the one thing that, if you forget everything else, still keeps you safe. AI never changes your document silently. Propose, preview, decide, undo. Everything else is detail.
Jessica: I'm Jessica.
Liam: I'm Liam. Thanks for listening. We will see you at episode thirty-eight.
Jessica: The QUILL Cast is a fifty-four-episode audio course on QUILL. Subscribe wherever you listen, and the show notes have every link, every keybinding, and every transcript.