diff --git a/.claude/skills/questdb-docs-writer/SKILL.md b/.claude/skills/questdb-docs-writer/SKILL.md
index 1b3c42ab9..728d145c3 100644
--- a/.claude/skills/questdb-docs-writer/SKILL.md
+++ b/.claude/skills/questdb-docs-writer/SKILL.md
@@ -710,6 +710,24 @@ in `src/theme/MDXComponents.js`). Store screenshots as `.webp` in
 cwebp -lossless -z 9 -m 6 -mt input.png -o output.webp
 ```
 
+Commit only the WebP. The repo's lint-staged hook quantizes staged PNG and
+JPEG files lossily (pngquant, mozjpeg), and leaves WebP alone.
+
+#### Capture at 2x
+
+A screenshot shown at `width={800}` needs to be at least 1600 px wide to stay
+sharp on a retina screen, captured as PNG and kept lossless until the WebP.
+Screenshots from the Claude in Chrome `screenshot` action are downscaled 1x
+JPEGs and look blurry; so does a wide window shrunk into the column. Capture
+Web Console and other UI screenshots with headless Chrome at
+`deviceScaleFactor: 2` instead. Read
+[references/ui-screenshots/README.md](references/ui-screenshots/README.md)
+before taking any: it covers loading notebooks from the user's browser into a
+throwaway profile, running every cell, cleaning the frame, showing "MCP
+connected" without opening the user's browser, and the encode check. The
+driver script `references/ui-screenshots/cdp-shot.mjs` needs only Node 22 and
+Chrome.
+
 ## SQL parser sync check
 
 When documenting new SQL features, check whether the `questdb/sql-parser`
diff --git a/.claude/skills/questdb-docs-writer/references/ui-screenshots/README.md b/.claude/skills/questdb-docs-writer/references/ui-screenshots/README.md
new file mode 100644
index 000000000..d507a6a5a
--- /dev/null
+++ b/.claude/skills/questdb-docs-writer/references/ui-screenshots/README.md
@@ -0,0 +1,98 @@
+# Sharp UI screenshots (Web Console and other browser UIs)
+
+How to capture screenshots of a browser UI that stay sharp on a retina
+screen, for blog posts and docs alike. The driver script `cdp-shot.mjs` sits
+next to this file.
+
+## Why screenshots come out blurry
+
+Two things make a screenshot blurry, and either one is enough:
+
+1. **It was captured at 1x.** A browser-automation screenshot (the Claude in
+   Chrome `screenshot` action, or a window on a non-retina monitor) is 1x, and
+   the extension also downscales and saves it as JPEG. Text then gets
+   upscaled on every retina screen. Converting that JPEG to WebP adds a
+   second lossy step.
+2. **It shows too wide a window.** A 1500 CSS px console shown in a column
+   880 px wide is shrunk to 58%, so its text ends up at about 7 px.
+
+Rule of thumb: the image must be at least **twice the width it is displayed
+at** (the blog column is 880 CSS px and docs screenshots usually render with
+`width={800}`, so 1600 to 1760 px or more), captured as PNG, and kept lossless
+until the final WebP.
+
+## What does not work
+
+- **The Claude in Chrome screenshot action.** It saves downscaled 1x JPEGs.
+  While the extension controls a tab, Chrome also shows a "started debugging
+  this browser" bar and an overlay on that window, so an OS screenshot of the
+  user's window includes them.
+- **macOS `screencapture` from the shell.** It needs Screen Recording
+  permission for the terminal. Don't change that setting; use headless
+  Chrome instead.
+- **Asking the user to move windows or take the shots.** Headless Chrome needs
+  nothing from them.
+
+## The workflow: headless Chrome at 2x over CDP
+
+1. **Start a headless Chrome with a throwaway profile** (command at the top of
+   `cdp-shot.mjs`). Device metrics are overridden to `deviceScaleFactor: 2`,
+   so a 1512 CSS px viewport captures at 3024 px, independent of any monitor.
+
+2. **Open the page** with `node cdp-shot.mjs nav <url>`.
+
+3. **Bring over browser-local state.** Web Console notebooks live in the
+   browser's IndexedDB (database `web-console`, store `buffers`), per origin.
+   A fresh profile has none. Export the records from the browser that has
+   them, by reading `buffers` with `getAll()` in a tab on the same origin
+   (records with a `notebookViewState` are notebooks). Then, at the same
+   origin in headless Chrome, run `node cdp-shot.mjs inject export.json` and
+   navigate again so the console picks them up. Exported notebooks are worth
+   keeping anyway: they are the reproducible record of what an agent built.
+
+4. **Open the notebook and run it.** Query results are not part of the
+   record, so table cells show their SQL until they run. The notebook's
+   refresh button only re-runs charts. Run every cell:
+
+   ```bash
+   node cdp-shot.mjs eval "(async () => {
+     const bs = [...document.querySelectorAll('button[aria-label=\"Run cell\"]')];
+     for (const b of bs) { b.click(); await new Promise(r => setTimeout(r, 4000)) }
+     return bs.length })()"
+   ```
+
+5. **Clean the frame.** Dismiss what isn't the subject: the
+   `vm.max_map_count` warning banner, the "Work with your coding agent via
+   MCP" promo card, and the schema side panel. Use real clicks
+   (`cdp-shot.mjs click x y`); a DOM `.click()` on a console tab doesn't
+   switch tabs. Dismissing a banner moves everything below it, so take a
+   fresh shot before computing the next click position.
+
+6. **Frame by element, not by guesswork.** Measure panels with
+   `getBoundingClientRect()`, scroll the panel into view with
+   `scrollIntoView({ block: 'start' })`, then nudge the scrolling container
+   back a little for margin. Raise `VH` until the panels you want fit, and
+   crop with the `shot` clip (`shot out.png 0 0 1512 940`). Keep the same `VW`
+   and `VH` for the measuring step and the shot, because each command
+   re-applies the viewport.
+
+7. **Showing "MCP connected" in the status bar.** Start an MCP session for
+   that console origin and call `get_pairing_credentials` with
+   `auto_open_browser: false`; otherwise the server opens the deep link in the
+   user's default browser. Open the deep link in headless Chrome, click
+   **Connect** with the permission the story needs, and stop the session after
+   the shot.
+
+8. **Encode and check.** Encode UI screenshots as lossless WebP, which for flat
+   UI is both the best quality and the smallest file (q92 lossy came out
+   about twice the size):
+
+   ```bash
+   cwebp -lossless -z 9 -mt shot.png -o image.webp
+   sips -g pixelWidth -g pixelHeight image.webp   # expect >= 2x display width
+   ```
+
+   Then read the image back and check that the text is crisp, the panels
+   have finished loading, and nothing is half-rendered.
+
+Keep the source PNGs out of the repo; commit only the WebP.
diff --git a/.claude/skills/questdb-docs-writer/references/ui-screenshots/cdp-shot.mjs b/.claude/skills/questdb-docs-writer/references/ui-screenshots/cdp-shot.mjs
new file mode 100644
index 000000000..4c19d316a
--- /dev/null
+++ b/.claude/skills/questdb-docs-writer/references/ui-screenshots/cdp-shot.mjs
@@ -0,0 +1,108 @@
+// Drive a headless Chrome over CDP to capture sharp 2x UI screenshots.
+// No npm dependencies: Node 22+ (global WebSocket and fetch).
+//
+// Start Chrome first, with a throwaway profile:
+//   "/Applications/Google Chrome.app/Contents/MacOS/Google Chrome" \
+//     --headless=new --remote-debugging-port=9333 \
+//     --user-data-dir="$TMPDIR/shot-profile" --hide-scrollbars \
+//     --no-first-run --no-default-browser-check about:blank
+//
+// Commands (VW/VH env vars set the viewport in CSS px, default 1512x860):
+//   node cdp-shot.mjs nav <url>                 navigate, wait WAIT ms (5000)
+//   node cdp-shot.mjs eval <js-expression>      print the JSON result
+//   node cdp-shot.mjs evalfile <file.js>
+//   node cdp-shot.mjs click <x> <y> [<x> <y>..] real mouse clicks, CSS px
+//   node cdp-shot.mjs inject <export.json>      load Web Console notebooks
+//   node cdp-shot.mjs shot <out.png> [x y w h]  PNG at 2x, optional CSS clip
+//
+// Every command reconnects and re-applies the 2x device metrics, so keep VW
+// and VH the same between a layout step and the shot that follows it.
+import fs from "node:fs"
+
+const PORT = process.env.CDP_PORT || 9333
+const [cmd, ...args] = process.argv.slice(2)
+const base = `http://127.0.0.1:${PORT}`
+const targets = await (await fetch(`${base}/json/list`)).json()
+let page = targets.find((t) => t.type === "page")
+if (!page) page = await (await fetch(`${base}/json/new?about:blank`, { method: "PUT" })).json()
+
+const ws = new WebSocket(page.webSocketDebuggerUrl)
+await new Promise((r) => ws.addEventListener("open", r, { once: true }))
+let seq = 0
+const pending = new Map()
+ws.addEventListener("message", (e) => {
+  const m = JSON.parse(e.data)
+  if (m.id && pending.has(m.id)) {
+    pending.get(m.id)(m)
+    pending.delete(m.id)
+  }
+})
+const send = (method, params = {}) =>
+  new Promise((resolve, reject) => {
+    const id = ++seq
+    pending.set(id, (m) => (m.error ? reject(new Error(JSON.stringify(m.error))) : resolve(m.result)))
+    ws.send(JSON.stringify({ id, method, params }))
+  })
+const sleep = (ms) => new Promise((r) => setTimeout(r, ms))
+const evaluate = async (expr) => {
+  const r = await send("Runtime.evaluate", { expression: expr, awaitPromise: true, returnByValue: true })
+  if (r.exceptionDetails) throw new Error(r.exceptionDetails.exception?.description || r.exceptionDetails.text)
+  return r.result.value
+}
+
+await send("Emulation.setDeviceMetricsOverride", {
+  width: Number(process.env.VW || 1512),
+  height: Number(process.env.VH || 860),
+  deviceScaleFactor: 2,
+  mobile: false,
+})
+
+if (cmd === "nav") {
+  await send("Page.enable")
+  await send("Page.navigate", { url: args[0] })
+  await sleep(Number(process.env.WAIT || 5000))
+  console.log(await evaluate("document.title + ' ' + location.href"))
+} else if (cmd === "eval") {
+  console.log(JSON.stringify(await evaluate(args.join(" ")), null, 1))
+} else if (cmd === "evalfile") {
+  console.log(JSON.stringify(await evaluate(fs.readFileSync(args[0], "utf8")), null, 1))
+} else if (cmd === "click") {
+  for (let i = 0; i < args.length; i += 2) {
+    for (const type of ["mousePressed", "mouseReleased"])
+      await send("Input.dispatchMouseEvent", { type, x: +args[i], y: +args[i + 1], button: "left", clickCount: 1 })
+    await sleep(Number(process.env.WAIT || 800))
+  }
+  console.log("clicked")
+} else if (cmd === "inject") {
+  // Accepts an array of `buffers` records, or { notebooks: [...] }.
+  const data = JSON.parse(fs.readFileSync(args[0], "utf8"))
+  const recs = Array.isArray(data) ? data : data.notebooks
+  console.log(
+    JSON.stringify(
+      await evaluate(`(async () => {
+        const recs = ${JSON.stringify(recs)};
+        const db = await new Promise((res, rej) => {
+          const r = indexedDB.open('web-console');
+          r.onsuccess = () => res(r.result);
+          r.onerror = () => rej(r.error);
+        });
+        const tx = db.transaction('buffers', 'readwrite');
+        for (const rec of recs) tx.objectStore('buffers').put(rec);
+        await new Promise((res, rej) => { tx.oncomplete = res; tx.onerror = () => rej(tx.error); });
+        db.close();
+        return recs.map((r) => r.label);
+      })()`),
+    ),
+  )
+} else if (cmd === "shot") {
+  const [out, x, y, w, h] = args
+  const params = { format: "png" }
+  if (w) params.clip = { x: +x, y: +y, width: +w, height: +h, scale: 1 }
+  const r = await send("Page.captureScreenshot", params)
+  fs.writeFileSync(out, Buffer.from(r.data, "base64"))
+  console.log("saved", out)
+} else {
+  console.error("unknown command:", cmd)
+  process.exitCode = 1
+}
+ws.close()