ZeroWidth

Lists the workspace's Napkin docs (quick markdown documents — drafts, notes, working text; the LINEAR surface, distinct from boards which are drawing canvases) newest first, with excerpts. Archived docs are hidden unless `includeArchived`; `q` narrows by title. Fetch one with napkin_docs_get. When referring the user to a doc in your reply, put `[[doc:ID]]` alone on its own line — it renders as a clickable card. NEVER cite a doc as [[board:ID]]; boards are sketches.

Edit a Napkin doc. For a change to existing prose use `edits` — a list of exact find/replace pairs, which is the SAFEST option and the one to reach for by default: it leaves everything you did not target untouched, and it fails loudly rather than clobbering a doc somebody else is typing in. `append` adds a section at the end. `body` replaces the WHOLE document and should be a last resort. Also retitle, describe, set visibility, or `archived` (true takes it out of the gallery, false brings it back). Read the doc with napkin_docs_get first — `edits` match the current text exactly. Doc ids come from napkin_docs_list.

Draws shapes, strokes, and labels on a board — diagram on a sketch, annotate a deck slide — and, via `updates` / `deletes`, changes or removes shapes that are already there (ids from napkin_boards_get; the fix for something you drew wrong). For NEW elements never compute absolute board coordinates: with `slide`, coordinates are slide-local — (0, 0) top-left to (960, 540) bottom-right of that slide. Without `slide` (sketches), (0, 0) is the top-left of the existing content — the same area `napkin_boards_view` renders, so place by what you saw there; on an empty board just start at (0, 0). Example — circle a slide's title and margin-note it: {"boardId": "…", "slide": 2, "elements": [ {"type": "ellipse", "x": 60, "y": 40, "w": 400, "h": 90, "color": "#dc2626"}, {"type": "arrow", "x": 560, "y": 140, "x2": 470, "y2": 90, "color": "#dc2626"}, {"type": "text", "x": 575, "y": 130, "w": 260, "text": "tighten this claim", "color": "#dc2626"} ]} After drawing, cite the board with `[[board:ID]]` alone on its own line.

Create an interface — a small screen of its own the user can open and send to people. Use when someone asks for a custom chat UI, a little tool, or a prototype of how something would feel in their product. It starts with a working index.html you then edit. It can reach NOTHING in the workspace until a flow is granted with napkin_interfaces_grant. Cite it in your reply as `[[interface:ID]]` alone on its own line — it renders as a card the user can click. Do NOT paste an image URL or try to embed the picture yourself; that renders as a broken image. An interface is plain HTML, CSS and JavaScript in one or more files, served from its own origin. Put styles in a <style> block. LIBRARIES: there is no build step and no CDN, and a <script src> to any other site is blocked. A fixed shelf of libraries is served from the interface's own origin instead; load them with exactly these tags (the versions are pinned, and any other /lib/ path is not found): PREACT + HTM, for a page with several screens or a lot of state — components and hooks with no build: <script src="/lib/[email protected]/preact.min.umd.js"></script> <script src="/lib/[email protected]/hooks.umd.js"></script> <script src="/lib/[email protected]/htm.umd.js"></script> then var html = htm.bind(preact.h); var useState = preactHooks.useState; function App(){ var s = useState(0); return html`<button onClick=${function(){ s[1](s[0] + 1) }}>${s[0]}</button>` } preact.render(html`<${App}/>`, document.getElementById("app")). Templates are tagged strings (html`…`), not JSX. CHART.JS, for a chart: <script src="/lib/[email protected]/chart.umd.min.js"></script> then new Chart(canvasElement, { type: "bar" | "line" | "doughnut" | …, data: { labels: [...], datasets: [{ label, data: [...] }] }, options: {} }). Color the datasets from the brand variables, read with getComputedStyle(document.documentElement).getPropertyValue("--brand-accent"). LUCIDE, for icons: <script src="/lib/[email protected]/lucide.min.js"></script>. In plain markup, write <i data-lucide="calendar"></i> and call lucide.createIcons() once it's on the page (again after adding more). In a Preact page DON'T — createIcons swaps elements Preact owns; draw the icon from its shapes instead: lucide.icons.Calendar is a list of [tag, attributes], so render <svg viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2"> with preact.h(tag, attributes) for each (the app starter has an Icon component that does this). Names are lucide.dev's: kebab-case in markup, PascalCase in lucide.icons. A one-screen page needs none of these: plain HTML, CSS and JavaScript is still the default. React, Tailwind and every other library are not available. index.html is the entry point and must exist. THE BRAND: every interface carries the workspace brand as files — link it with <link rel="stylesheet" href="brand.css"> and style with its variables: var(--brand-ink), var(--brand-muted), var(--brand-background), var(--brand-surface), var(--brand-accent), var(--brand-accent-2), var(--brand-color-<colour name>) (napkin_interfaces_view lists the exact names this brand defines), var(--brand-font-heading), var(--brand-font-body), var(--brand-radius). Never hard-code the brand's colours or font names. brand.css already loads the brand's own font files; web fonts (Google Fonts or any other) are BLOCKED, so linking one leaves the page in a system font. The brand's logos are files under brand/ (brand.css lists them at the top): <img src="brand/<name>.svg">. Use the logo rather than typing the name, and real icons (inline SVG) rather than emoji. Read the brand with napkin_brand_get first, and say only what the brand kit or the user tells you. PICTURES: a picture from anywhere on the web is BLOCKED and draws as a broken image, so never use an external image URL. Show a picture from the workspace's files with <img src="file:<fileId>" alt="…"> (in CSS, url(file:<fileId>)) — PNG, JPEG, GIF or WebP, up to 5 MB. Write the reference literally; a file id assembled in JavaScript is not found. When the user gives you a picture in chat, save it with workspace_files_save_from_chat and use the file id it returns. A picture that's only on a website has to be attached in chat first, or uploaded in the interface's Files panel. LOOK before you hand it over: napkin_interfaces_view draws the page as it is now. Fix anything that doesn't look like the brand and look again. Load the client with <script src="/zerowidth.js"></script>. Then zw.ready() resolves with the context { viewer, grants, flows, tables, metrics, tasks }; ctx.flows lists the flow grants, each with its `id`, `name` and `variables`. zw.flows.run(flow.id, input) runs one. Call it inside zw.ready().then(...), and handle the rejection — it rejects only when the page isn't open in ZeroWidth, so show "Open this from ZeroWidth" there instead of a dead button. A run can take up to two minutes; show it working (stream it, below), and show the error's message when one fails — err.message is a sentence written for the person using the page. Input is {kind:"chat", messages:[{role:"user", content:"…"}]} or {kind:"form", values:{…}} — the same shapes the public API takes. zw.replyText(result) pulls the assistant text out of a chat result. SOURCES: when the flow searches a knowledge base, zw.sources(result) lists the documents the run actually read, as [{ name, documentId }], once each. Show those under the answer for "where did this come from" — never ask the model to write a sources line and parse it out of the text. The flow needs the knowledge base attached BEFORE it is published (workbench_flows_attach_knowledge_base, then workbench_flows_publish, then re-grant the new version here); the interface runs the published version it was granted. FEEDBACK: to let someone rate an answer or flag a wrong one, call zw.feedback.send(result, { rating: "up" | "down", correction, comment }) with the result zw.flows.run returned. It goes to Caliper, into the flow's feedback dataset, with the question and the answer taken from the run itself; `correction` is what the answer should have been. Needs no extra grant and no table — never make a table to log flagged answers. It returns a promise: say it was sent only once it resolves, and show the error's message if it rejects. Tell the user flagged answers are in Caliper, in the dataset named after the flow, under Datasets. VARIABLES: a flow grant lists `variables` — the values the flow reads besides the conversation, usually the {{key}} tokens in its prompt (who the user is, their plan, the page they are on). Send them on the same chat run: {kind:"chat", messages:[…], values:{<name>: value}}, using each variable's exact `name` and its `type` (string, number, boolean, array, object). A variable left out reaches the prompt empty. Read them from napkin_interfaces_get while you write the page, or at run time from ctx.flows[i].variables. A SHIM is not a flow and takes a different call: zw.shims.run(shimId, text), which decides on the viewer's own device with no network and no cost. It resolves with the same envelope a flow does, so read the decision with zw.decision(result) — NOT result.decision, which is undefined and makes an interface show one answer for every input. The decision is { answer, confidence, familiarity, action, probs, gates }; branch on `action`, the shim's own call about whether it was sure enough. Calling zw.flows.run with a shim id is refused. ctx.grants tells you which you have: each entry carries a `kind` of "flow" or "shim" alongside its id and name. DATA: an interface can be granted a sheet TABLE or a Ledger METRIC with napkin_interfaces_grant_data. ctx.tables lists each table with its `fields` (name, type, options) and `access` ("read", "add" or "edit"); ctx.metrics lists each metric with its unit and `access` ("read" or "record"). Records travel keyed by field NAME, choices by option NAME. zw.tables.query(table, { where: [{ field, op, value }], sort: [{ field, dir }], limit }) resolves { records: [{ id, fields }], total }; zw.tables.add(table, { Field: value }) or a list adds records; zw.tables.update(table, id, fields) and zw.tables.remove(table, ids) need "edit". zw.metrics.read(metric, { from, to }) resolves { metric, readings: [{ value, at, note }] }; zw.metrics.record(metric, value, { note }) needs "record". Name a table or metric by its id or its name. Through a share link a guest can read and add, never change or delete — so an intake form works when shared, a triage board needs a signed-in viewer. A granted Workbench TASK (ctx.tasks; each says whether it `takesInput`) starts with zw.tasks.start(task, input), which resolves { runId, status } at once — the run carries on after the page closes — and zw.tasks.get(task, runId) reads { status, output } for runs this page started. Running a shim on every keystroke is fine — it costs nothing and there is no rate limit. Debounce ~150ms and COALESCE: remember the latest text and run it when the current call finishes, so the answer matches what is on screen. Clear any in-flight guard on failure as well as success, or one call that doesn't come back wedges the interface. `d.action` is "act" | "suggest" | "refuse" — those three strings, nothing else. Branch on it rather than on a confidence threshold you invent; "refuse" means the shim doesn't recognise the input well enough to answer, so say so rather than showing a low-confidence guess. If you run a requestAnimationFrame loop, remove CSS transitions from any property it writes — the two fight and the property looks frozen. A flow answers in markdown: render it with el.replaceChildren(zw.markdown(text)). It builds DOM nodes, so model output is never treated as markup. STREAM chat answers: zw.flows.run(id, input, { onEvent: fn }) delivers the text token by token, and a flow that takes several seconds reads as broken without it. Use zw.textDelta(event) for each chunk — it returns null for anything that isn't text, so pass it every event — and accumulate. The promise still …

Read one interface: its name, the flows it may run (with the published version each is pinned to), and the contents of its files. READ BEFORE EDITING — napkin_interfaces_write matches text exactly against what is there now. Pass a path to read one file when the interface has several. An interface is plain HTML, CSS and JavaScript in one or more files, served from its own origin. Put styles in a <style> block. LIBRARIES: there is no build step and no CDN, and a <script src> to any other site is blocked. A fixed shelf of libraries is served from the interface's own origin instead; load them with exactly these tags (the versions are pinned, and any other /lib/ path is not found): PREACT + HTM, for a page with several screens or a lot of state — components and hooks with no build: <script src="/lib/[email protected]/preact.min.umd.js"></script> <script src="/lib/[email protected]/hooks.umd.js"></script> <script src="/lib/[email protected]/htm.umd.js"></script> then var html = htm.bind(preact.h); var useState = preactHooks.useState; function App(){ var s = useState(0); return html`<button onClick=${function(){ s[1](s[0] + 1) }}>${s[0]}</button>` } preact.render(html`<${App}/>`, document.getElementById("app")). Templates are tagged strings (html`…`), not JSX. CHART.JS, for a chart: <script src="/lib/[email protected]/chart.umd.min.js"></script> then new Chart(canvasElement, { type: "bar" | "line" | "doughnut" | …, data: { labels: [...], datasets: [{ label, data: [...] }] }, options: {} }). Color the datasets from the brand variables, read with getComputedStyle(document.documentElement).getPropertyValue("--brand-accent"). LUCIDE, for icons: <script src="/lib/[email protected]/lucide.min.js"></script>. In plain markup, write <i data-lucide="calendar"></i> and call lucide.createIcons() once it's on the page (again after adding more). In a Preact page DON'T — createIcons swaps elements Preact owns; draw the icon from its shapes instead: lucide.icons.Calendar is a list of [tag, attributes], so render <svg viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2"> with preact.h(tag, attributes) for each (the app starter has an Icon component that does this). Names are lucide.dev's: kebab-case in markup, PascalCase in lucide.icons. A one-screen page needs none of these: plain HTML, CSS and JavaScript is still the default. React, Tailwind and every other library are not available. index.html is the entry point and must exist. THE BRAND: every interface carries the workspace brand as files — link it with <link rel="stylesheet" href="brand.css"> and style with its variables: var(--brand-ink), var(--brand-muted), var(--brand-background), var(--brand-surface), var(--brand-accent), var(--brand-accent-2), var(--brand-color-<colour name>) (napkin_interfaces_view lists the exact names this brand defines), var(--brand-font-heading), var(--brand-font-body), var(--brand-radius). Never hard-code the brand's colours or font names. brand.css already loads the brand's own font files; web fonts (Google Fonts or any other) are BLOCKED, so linking one leaves the page in a system font. The brand's logos are files under brand/ (brand.css lists them at the top): <img src="brand/<name>.svg">. Use the logo rather than typing the name, and real icons (inline SVG) rather than emoji. Read the brand with napkin_brand_get first, and say only what the brand kit or the user tells you. PICTURES: a picture from anywhere on the web is BLOCKED and draws as a broken image, so never use an external image URL. Show a picture from the workspace's files with <img src="file:<fileId>" alt="…"> (in CSS, url(file:<fileId>)) — PNG, JPEG, GIF or WebP, up to 5 MB. Write the reference literally; a file id assembled in JavaScript is not found. When the user gives you a picture in chat, save it with workspace_files_save_from_chat and use the file id it returns. A picture that's only on a website has to be attached in chat first, or uploaded in the interface's Files panel. LOOK before you hand it over: napkin_interfaces_view draws the page as it is now. Fix anything that doesn't look like the brand and look again. Load the client with <script src="/zerowidth.js"></script>. Then zw.ready() resolves with the context { viewer, grants, flows, tables, metrics, tasks }; ctx.flows lists the flow grants, each with its `id`, `name` and `variables`. zw.flows.run(flow.id, input) runs one. Call it inside zw.ready().then(...), and handle the rejection — it rejects only when the page isn't open in ZeroWidth, so show "Open this from ZeroWidth" there instead of a dead button. A run can take up to two minutes; show it working (stream it, below), and show the error's message when one fails — err.message is a sentence written for the person using the page. Input is {kind:"chat", messages:[{role:"user", content:"…"}]} or {kind:"form", values:{…}} — the same shapes the public API takes. zw.replyText(result) pulls the assistant text out of a chat result. SOURCES: when the flow searches a knowledge base, zw.sources(result) lists the documents the run actually read, as [{ name, documentId }], once each. Show those under the answer for "where did this come from" — never ask the model to write a sources line and parse it out of the text. The flow needs the knowledge base attached BEFORE it is published (workbench_flows_attach_knowledge_base, then workbench_flows_publish, then re-grant the new version here); the interface runs the published version it was granted. FEEDBACK: to let someone rate an answer or flag a wrong one, call zw.feedback.send(result, { rating: "up" | "down", correction, comment }) with the result zw.flows.run returned. It goes to Caliper, into the flow's feedback dataset, with the question and the answer taken from the run itself; `correction` is what the answer should have been. Needs no extra grant and no table — never make a table to log flagged answers. It returns a promise: say it was sent only once it resolves, and show the error's message if it rejects. Tell the user flagged answers are in Caliper, in the dataset named after the flow, under Datasets. VARIABLES: a flow grant lists `variables` — the values the flow reads besides the conversation, usually the {{key}} tokens in its prompt (who the user is, their plan, the page they are on). Send them on the same chat run: {kind:"chat", messages:[…], values:{<name>: value}}, using each variable's exact `name` and its `type` (string, number, boolean, array, object). A variable left out reaches the prompt empty. Read them from napkin_interfaces_get while you write the page, or at run time from ctx.flows[i].variables. A SHIM is not a flow and takes a different call: zw.shims.run(shimId, text), which decides on the viewer's own device with no network and no cost. It resolves with the same envelope a flow does, so read the decision with zw.decision(result) — NOT result.decision, which is undefined and makes an interface show one answer for every input. The decision is { answer, confidence, familiarity, action, probs, gates }; branch on `action`, the shim's own call about whether it was sure enough. Calling zw.flows.run with a shim id is refused. ctx.grants tells you which you have: each entry carries a `kind` of "flow" or "shim" alongside its id and name. DATA: an interface can be granted a sheet TABLE or a Ledger METRIC with napkin_interfaces_grant_data. ctx.tables lists each table with its `fields` (name, type, options) and `access` ("read", "add" or "edit"); ctx.metrics lists each metric with its unit and `access` ("read" or "record"). Records travel keyed by field NAME, choices by option NAME. zw.tables.query(table, { where: [{ field, op, value }], sort: [{ field, dir }], limit }) resolves { records: [{ id, fields }], total }; zw.tables.add(table, { Field: value }) or a list adds records; zw.tables.update(table, id, fields) and zw.tables.remove(table, ids) need "edit". zw.metrics.read(metric, { from, to }) resolves { metric, readings: [{ value, at, note }] }; zw.metrics.record(metric, value, { note }) needs "record". Name a table or metric by its id or its name. Through a share link a guest can read and add, never change or delete — so an intake form works when shared, a triage board needs a signed-in viewer. A granted Workbench TASK (ctx.tasks; each says whether it `takesInput`) starts with zw.tasks.start(task, input), which resolves { runId, status } at once — the run carries on after the page closes — and zw.tasks.get(task, runId) reads { status, output } for runs this page started. Running a shim on every keystroke is fine — it costs nothing and there is no rate limit. Debounce ~150ms and COALESCE: remember the latest text and run it when the current call finishes, so the answer matches what is on screen. Clear any in-flight guard on failure as well as success, or one call that doesn't come back wedges the interface. `d.action` is "act" | "suggest" | "refuse" — those three strings, nothing else. Branch on it rather than on a confidence threshold you invent; "refuse" means the shim doesn't recognise the input well enough to answer, so say so rather than showing a low-confidence guess. If you run a requestAnimationFrame loop, remove CSS transitions from any property it writes — the two fight and the property looks frozen. A flow answers in markdown: render it with el.replaceChildren(zw.markdown(text)). It builds DOM nodes, so model output is never treated as markup. STREAM chat answers: zw.flows.run(id, input, { onEvent: fn }) delivers the text token by token, and a flow that takes several seconds reads as broken without it. Use zw.textDelta(event) for each chunk — it returns null for anything that isn't text, so pass it every event — and accumulate. The promise still settles at the end with the whole result; take the final text from there. onEvent also sees node_start / node_complete / node_error if you want to name the step. Don't name a top-level variable `history`, `name`, `status`, `length`, `origin` or `top`: those are already…

Set what an interface may run — flows and shims, each pinned to a published version. This is the interface's ONLY reach into the workspace, and it replaces the whole list, so include everything it should keep. Ids and published versions come from workbench_flows_list / workbench_shim_list and their revisions. Ask the user before widening this.

Set the sheet tables, Ledger metrics and Workbench Tasks an interface may use (ADR 0085, 0086), replacing the whole list. Tables: access read, add (read and add records) or edit (also change and delete); optionally limit to some columns by name. Metrics: read or record. Tasks: the page may start them. The page then calls zw.tables.query / add / update / remove, zw.metrics.read / record and zw.tasks.start / get. Anyone who can open the interface, including people it's shared with, gets this access (a shared link can only read and add), so ask the user before widening it.

List the interfaces in a workspace — small self-contained screens (a custom chat, a control panel, a prototype) that run on their own origin and can call the flows they've been granted. Returns ids, names and how many flows each may run.

Save where an interface is now as a numbered version. Share links point at a published version, so what someone was sent doesn't change while work continues. Do this when the user says it's ready to show people, not after every edit.

See what an interface LOOKS like and DOES right now: the page RUNS from its current files under the same rules as the live page (no web fonts, no remote pictures), and you get a picture plus anything that went wrong (script errors, failed promises, blocked requests, console errors). Use it after every write and before saying an interface is finished: reading the code tells you what you wrote, not what it renders. `size: "phone"` looks at it 390px wide — use it for anything meant for a phone, like a mobile app mockup. `steps` drive the page before and between pictures, so you can check every screen and that buttons work: `{click: "Plan"}` (by visible text or label, or a CSS selector like "#nav-plan"), `{type: {into: "textarea", text: "hi"}}`, `{press: "Enter"}`, `{wait: 500}`, and `{shot: "Plan screen"}` for a picture. With no steps you get one picture of the first screen. Check each tab of an app in ONE call: shot, click a tab, shot, click the next, shot. Flows don't run here: a flow call gets a placeholder reply, so you can see the chat send and answer without spending anything; tables and metrics read empty. Look for text that's too small or runs together, stray system fonts, emoji standing in for icons, and anything that doesn't look like the brand.

Write one file of an interface. Use `edits` — exact find/replace pairs — which is the SAFEST option and the one to reach for by default: it leaves everything you did not target untouched, and it fails loudly rather than clobbering a file somebody is editing. `body` replaces the WHOLE file and is for a new file or a deliberate rewrite. ONE CALL HOLDS ABOUT 15,000 CHARACTERS of file: a longer `body` is cut off mid-write and fails. Write a big file in parts instead — the first part with `body`, then each further part with `body` plus `append: true`, in order — and split an app across files (index.html, style.css, app.js) so each stays readable. `deleteFile` removes one (index.html can't be removed). Read with napkin_interfaces_get first — `edits` match the current text exactly. Also rename, describe, set visibility, or archive. An interface is plain HTML, CSS and JavaScript in one or more files, served from its own origin. Put styles in a <style> block. LIBRARIES: there is no build step and no CDN, and a <script src> to any other site is blocked. A fixed shelf of libraries is served from the interface's own origin instead; load them with exactly these tags (the versions are pinned, and any other /lib/ path is not found): PREACT + HTM, for a page with several screens or a lot of state — components and hooks with no build: <script src="/lib/[email protected]/preact.min.umd.js"></script> <script src="/lib/[email protected]/hooks.umd.js"></script> <script src="/lib/[email protected]/htm.umd.js"></script> then var html = htm.bind(preact.h); var useState = preactHooks.useState; function App(){ var s = useState(0); return html`<button onClick=${function(){ s[1](s[0] + 1) }}>${s[0]}</button>` } preact.render(html`<${App}/>`, document.getElementById("app")). Templates are tagged strings (html`…`), not JSX. CHART.JS, for a chart: <script src="/lib/[email protected]/chart.umd.min.js"></script> then new Chart(canvasElement, { type: "bar" | "line" | "doughnut" | …, data: { labels: [...], datasets: [{ label, data: [...] }] }, options: {} }). Color the datasets from the brand variables, read with getComputedStyle(document.documentElement).getPropertyValue("--brand-accent"). LUCIDE, for icons: <script src="/lib/[email protected]/lucide.min.js"></script>. In plain markup, write <i data-lucide="calendar"></i> and call lucide.createIcons() once it's on the page (again after adding more). In a Preact page DON'T — createIcons swaps elements Preact owns; draw the icon from its shapes instead: lucide.icons.Calendar is a list of [tag, attributes], so render <svg viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2"> with preact.h(tag, attributes) for each (the app starter has an Icon component that does this). Names are lucide.dev's: kebab-case in markup, PascalCase in lucide.icons. A one-screen page needs none of these: plain HTML, CSS and JavaScript is still the default. React, Tailwind and every other library are not available. index.html is the entry point and must exist. THE BRAND: every interface carries the workspace brand as files — link it with <link rel="stylesheet" href="brand.css"> and style with its variables: var(--brand-ink), var(--brand-muted), var(--brand-background), var(--brand-surface), var(--brand-accent), var(--brand-accent-2), var(--brand-color-<colour name>) (napkin_interfaces_view lists the exact names this brand defines), var(--brand-font-heading), var(--brand-font-body), var(--brand-radius). Never hard-code the brand's colours or font names. brand.css already loads the brand's own font files; web fonts (Google Fonts or any other) are BLOCKED, so linking one leaves the page in a system font. The brand's logos are files under brand/ (brand.css lists them at the top): <img src="brand/<name>.svg">. Use the logo rather than typing the name, and real icons (inline SVG) rather than emoji. Read the brand with napkin_brand_get first, and say only what the brand kit or the user tells you. PICTURES: a picture from anywhere on the web is BLOCKED and draws as a broken image, so never use an external image URL. Show a picture from the workspace's files with <img src="file:<fileId>" alt="…"> (in CSS, url(file:<fileId>)) — PNG, JPEG, GIF or WebP, up to 5 MB. Write the reference literally; a file id assembled in JavaScript is not found. When the user gives you a picture in chat, save it with workspace_files_save_from_chat and use the file id it returns. A picture that's only on a website has to be attached in chat first, or uploaded in the interface's Files panel. LOOK before you hand it over: napkin_interfaces_view draws the page as it is now. Fix anything that doesn't look like the brand and look again. Load the client with <script src="/zerowidth.js"></script>. Then zw.ready() resolves with the context { viewer, grants, flows, tables, metrics, tasks }; ctx.flows lists the flow grants, each with its `id`, `name` and `variables`. zw.flows.run(flow.id, input) runs one. Call it inside zw.ready().then(...), and handle the rejection — it rejects only when the page isn't open in ZeroWidth, so show "Open this from ZeroWidth" there instead of a dead button. A run can take up to two minutes; show it working (stream it, below), and show the error's message when one fails — err.message is a sentence written for the person using the page. Input is {kind:"chat", messages:[{role:"user", content:"…"}]} or {kind:"form", values:{…}} — the same shapes the public API takes. zw.replyText(result) pulls the assistant text out of a chat result. SOURCES: when the flow searches a knowledge base, zw.sources(result) lists the documents the run actually read, as [{ name, documentId }], once each. Show those under the answer for "where did this come from" — never ask the model to write a sources line and parse it out of the text. The flow needs the knowledge base attached BEFORE it is published (workbench_flows_attach_knowledge_base, then workbench_flows_publish, then re-grant the new version here); the interface runs the published version it was granted. FEEDBACK: to let someone rate an answer or flag a wrong one, call zw.feedback.send(result, { rating: "up" | "down", correction, comment }) with the result zw.flows.run returned. It goes to Caliper, into the flow's feedback dataset, with the question and the answer taken from the run itself; `correction` is what the answer should have been. Needs no extra grant and no table — never make a table to log flagged answers. It returns a promise: say it was sent only once it resolves, and show the error's message if it rejects. Tell the user flagged answers are in Caliper, in the dataset named after the flow, under Datasets. VARIABLES: a flow grant lists `variables` — the values the flow reads besides the conversation, usually the {{key}} tokens in its prompt (who the user is, their plan, the page they are on). Send them on the same chat run: {kind:"chat", messages:[…], values:{<name>: value}}, using each variable's exact `name` and its `type` (string, number, boolean, array, object). A variable left out reaches the prompt empty. Read them from napkin_interfaces_get while you write the page, or at run time from ctx.flows[i].variables. A SHIM is not a flow and takes a different call: zw.shims.run(shimId, text), which decides on the viewer's own device with no network and no cost. It resolves with the same envelope a flow does, so read the decision with zw.decision(result) — NOT result.decision, which is undefined and makes an interface show one answer for every input. The decision is { answer, confidence, familiarity, action, probs, gates }; branch on `action`, the shim's own call about whether it was sure enough. Calling zw.flows.run with a shim id is refused. ctx.grants tells you which you have: each entry carries a `kind` of "flow" or "shim" alongside its id and name. DATA: an interface can be granted a sheet TABLE or a Ledger METRIC with napkin_interfaces_grant_data. ctx.tables lists each table with its `fields` (name, type, options) and `access` ("read", "add" or "edit"); ctx.metrics lists each metric with its unit and `access` ("read" or "record"). Records travel keyed by field NAME, choices by option NAME. zw.tables.query(table, { where: [{ field, op, value }], sort: [{ field, dir }], limit }) resolves { records: [{ id, fields }], total }; zw.tables.add(table, { Field: value }) or a list adds records; zw.tables.update(table, id, fields) and zw.tables.remove(table, ids) need "edit". zw.metrics.read(metric, { from, to }) resolves { metric, readings: [{ value, at, note }] }; zw.metrics.record(metric, value, { note }) needs "record". Name a table or metric by its id or its name. Through a share link a guest can read and add, never change or delete — so an intake form works when shared, a triage board needs a signed-in viewer. A granted Workbench TASK (ctx.tasks; each says whether it `takesInput`) starts with zw.tasks.start(task, input), which resolves { runId, status } at once — the run carries on after the page closes — and zw.tasks.get(task, runId) reads { status, output } for runs this page started. Running a shim on every keystroke is fine — it costs nothing and there is no rate limit. Debounce ~150ms and COALESCE: remember the latest text and run it when the current call finishes, so the answer matches what is on screen. Clear any in-flight guard on failure as well as success, or one call that doesn't come back wedges the interface. `d.action` is "act" | "suggest" | "refuse" — those three strings, nothing else. Branch on it rather than on a confidence threshold you invent; "refuse" means the shim doesn't recognise the input well enough to answer, so say so rather than showing a low-confidence guess. If you run a requestAnimationFrame loop, remove CSS transitions from any property it writes — the two fight and the property looks frozen. A flow answers in markdown: render it with el.replaceChildren(zw.markdown(text)). It builds DOM nodes, so model output is never treated as markup. STREAM chat answers: zw.f…

Lists the slide layouts napkin_deck_compose can build from — the brand kit's own first, then the built-in ones — with when to use each and the slots it takes (`needs` must be filled; `takes` are optional). Every layout adapts to every size, fits its text, and uses the brand's colors, faces and logos; the logo fills itself in. Pick by what the slide has to say: one line (statement), a number (stat), a quote, a list (points, steps), an event, a carousel (cover, inner, closing), a picture (image-top, image-full, split, corner, product-shot), or a display ad (banner).

Adds a LIVE chart to a sheet — its SQL re-runs against the tabs on every view, so it never goes stale. Write the query with napkin_sheets_query first to confirm the shape (first column = x axis, numeric columns = series, or set x/series explicitly). Prefer aggregated queries (GROUP BY) — a chart of raw rows is rarely the answer.

Creates a Napkin sheet (multi-tab spreadsheet), optionally titled. Then populate it with napkin_sheets_set_cells (headers in row 1). Tell the user where it landed — the Sheets tab in Napkin.

Lists the workspace's Napkin sheets (multi-tab spreadsheets — the small-data surface) newest first with tab/chart counts. Archived sheets are hidden unless `includeArchived`; `q` narrows by title. Inspect one with napkin_sheets_schema, then query it with napkin_sheets_query.

Runs a SQLite SELECT over a sheet's tabs-as-tables (get the table/column names from napkin_sheets_schema first). Full SQLite dialect: WHERE, GROUP BY, ORDER BY, JOINs across tabs, aggregates. Results cap at 200 rows. This is THE way to answer questions about a sheet's data — never eyeball cells when a query can answer precisely.

Returns a sheet's tabs as SQL table definitions — CREATE TABLE scripts with row counts (each tab is a table; its header row names the columns; formula cells contribute computed values). Read this BEFORE writing a napkin_sheets_query so your SQL matches the real tables.

Sets cells in one tab of a sheet, by A1 address — values or formulas ('=SUM(B2:B9)'). Additive and surgical: only the addressed cells change. Check napkin_sheets_schema first so you know the tab names and where data ends; put headers in row 1 when creating a new region. To build a whole small table, write header cells + data cells in one call.

Edits a sheet's details: title, description, visibility, or `archived` (true takes it out of the gallery; false brings it back — the recovery move for a sheet created by mistake or no longer wanted). Cells are edited with napkin_sheets_set_cells, not here. Sheet ids come from napkin_sheets_list.

Renders one of a sheet's pinned charts against the LIVE data and returns the image so you can SEE it — the exact chart the user sees. The text part carries `imageUrl`; to show the user the chart inline in your reply, put a markdown image on its own line: ``. Use after napkin_sheets_add_chart to confirm the chart reads well, or whenever the user asks about a chart.

Appends one slide to an existing deck. Pick a layout (title | section | title-body | title-lead | two-column | three-column | comparison | statement | quote | closing | blank), give the title, and optionally `content` — markdown for the layout's remaining text areas, split by `---` lines (two-column/comparison take one block per column). Check the deck first with `napkin_deck_outline`.

Fills a slide's empty layout text areas by NAME — the names `napkin_deck_outline` surfaces as `[empty text stub: "…"]`. Pass `texts` mapping those exact names to content (markdown `- ` bullets welcome). Read the outline first to see which stubs a slide still has open.

Changes one slide as a whole: move it to another position in the deck (`moveTo`, 1-based — the other slides shift around it), hide it from the show without deleting it (`skipped`), show or hide its slide number (`pageNumber`), or replace its speaker notes (`notes`, markdown). Slide numbers come from napkin_deck_outline. To change what's ON the slide, use napkin_slide_fill or napkin_draw.

Renders a single slide of a Napkin deck to an image — exactly what present mode shows, cropped to the slide. Navigate by 1-based `slide` number in presentation order (get the map from `napkin_deck_outline` first; the result echoes `slideIndex`/`slideCount` so you can step through a deck slide by slide). Use this to check visual layout, drawings, and images that the text outline can't carry.

Adds columns to the end of a table. Existing records leave them empty.

Adds records, each a map of field name → value. Choices by option name (a new name becomes a new option); dates as YYYY-MM-DD; people by user id; checkboxes true/false. Every problem is reported at once and nothing is written until all records are valid.

Adds a saved view everyone sees: a grid (with where / sort), a board (lanes from a select or person column — groupBy), or a form people fill in to add records (formFields, formTitle, formIntro). A form's share link is made from the sheet's form view.

Makes a table tab — typed columns, one record per row, with grid, board and form views — in an existing sheet, or in a new sheet when sheetId is omitted. Use a table (not napkin_sheets_set_cells) for lists people keep: requests, vendors, bugs, contacts. The first column names each record. Column types: text, long_text, number, date, checkbox, select, multi_select, person, url, email.

Deletes records by id. People can undo this in the sheet only while it's open, so confirm with the person first.

A table tab's columns (name, type, choices), its views, and how many records it holds. Read this before writing records so field names and choices match.

Records from a table tab, keyed by field name, with optional where / sort / search / limit (max 500). For counts, sums and joins, napkin_sheets_query runs SQL over tables too (each table is an SQL table with an id column).

Changes some fields of records by id (from napkin_tables_query). Fields left out keep their values.

Creates a field from an origin idea — the seed of an exploration canvas. Counts against the workspace's field quota. May return `needs_confirmation` — tell the user what you're proposing and wait for approval.

Removes a field and its whole canvas (soft delete; frees the workspace's field quota). Use when the user is done with an exploration or one was created by mistake — studies attached to its nodes are not deleted. Field ids come from prism_fields_list. May return `needs_confirmation` — name the field and wait.

One field's full tree: every node (id, parent, axis, description, pinned, color, whether it has an image / desk research) plus the AI-clustered groups. The origin node has parentId=null; children sit along named axes. Use node ids with the expand / research / study tools.

Lists Prism fields (idea-exploration canvases) in the active workspace: name, origin idea, node count. Fetch one with `prism_fields_get` for its nodes.

Edits a field's name (null reverts to the origin idea), description, or visibility. Nodes are untouched. Field ids come from prism_fields_list. May return `needs_confirmation`.

Promotes an analysis theme (or, with no themeId, the analysis summary) into the workspace's Ledger as a belief, carrying the supporting verbatims as evidence with links back to the study. THE step that turns a finding into workspace memory — use when the user says an insight matters. May return `needs_confirmation`.

Creates a one-on-one AI-led interview on a study and returns the invite link to forward — the guest needs no account. The transcript stays on the study and joins the next analysis run (it never lands in Compass). focusPrompt briefs the interviewer on what to dig into; omitted, it explores the study's territory. May return `needs_confirmation`.

The study's one-on-one qual conversations: guest, status, focus, and — once completed — the structured memo (summary, key claims, tensions, verbatim quotes). Raw transcript included only while no memo exists. inviteUrl is the link to forward for interviews still waiting on their guest.

THE core Prism gesture: generate variations of a node along a named semantic axis ('more visceral', 'for the skeptic', 'stripped to essentials'…) in one of four directions. Runs generation against the workspace's inference credit. May return `needs_confirmation`.

Web-sourced secondary research on a node's idea — market context with citations, stored on the node. depth 'thorough' runs three angled passes (market, evidence, shifts) at ~3× the credit. `focus` steers what it goes after; without one it answers a generic brief off the node's own description, so pass it whenever the user has said what they actually want to know. Runs against the workspace's inference credit. May return `needs_confirmation`.

Sets pinned state and/or the color tag on a node. Pins are the cross-field shortlist. May return `needs_confirmation`.

Stops a queued, running or paused panel run. Answers already in stay. May return `needs_confirmation`.

What running a study against a panel would cost: people, model calls, and approximate dollars of workspace credit. Starts nothing. Tell the person the cost before calling prism_panels_run.

Status, progress, spend and the run's own analysis. For the numbers, call prism_studies_results with panelRunId. Synthetic answers are never people's: always say so when you report them.

Saves a panel from a spec (usually from prism_panels_draft) and draws its people from population records in the background, which takes a few seconds. Visible to the workspace unless visibility is PRIVATE. May return `needs_confirmation`.

Turns a plain description of an audience ("100 Midwest senior citizens") into a panel spec: filters on Census and General Social Survey fields, a size, and stated shares for traits with no data behind them (assumed). Saves nothing. Show the person the spec and assumedNotes, then pass the spec to prism_panels_create. poolSize is how many people in the population sample match.

One panel with its current version and the share of people by every field (assumed fields flagged). Use it to confirm a draw looks like the audience the person meant.

Every synthetic panel the person can see: name, current version (people, status), runs so far, and fit (how far it sat from people the last time a study had both). Reuse an existing panel before drafting a new one.