IMPORTANT - Path Resolution: This skill can be installed in different locations (plugin system, manual installation, global, or project-specific). Before executing any commands, determine the skill directory based on where you loaded this SKILL.md file, and use that path in all commands below. Replace $SKILL_DIR with the actual discovered path.

Demo

Records a scripted flow of the local app as video.mp4 (1440x900, 30fps, full X display with XFCE panel, Chromium window, and system cursor) plus events.json.

Input forms

The first word decides the form: init is a command; anything else is a demo name, path, or flow description.

1. Flow description, for example /demo record the checkout paying by card logged in as a customer. Pick a name, write .demos/<name>/script.mjs, record.
2. Existing demo, for example /demo checkout-card-as-customer, /demo .demos/checkout-card-as-customer, or /demo .demos/checkout-card-as-customer/script.mjs. Skip authoring and re-record as-is. If it fails loudly (the UI changed), fix the demo script, retry, and tell the user what changed in it.
3. init, i.e. /demo init: follow references/init.md.

Layout

One directory per demo, fixed file names, plus one optional config for all demos:

.demos/demo.config.mjs     settings for every demo (see Config)
.demos/checkout-card-as-customer/
  script.mjs     demo script
  video.mp4      last good recording
  events.json    telemetry of that recording
  failed.png     only after a failed run that reached the desktop
  failed.json    only after a failed run

A successful run replaces video.mp4 and events.json and removes failed.*. A failed run leaves the previous good video untouched. .demos/ is gitignored.

Names are descriptive kebab-case slugs of the specific flow, not the area: <area>-<flow>[-<variant>], e.g. checkout-card-as-customer, checkout-coupon-applied, admin-user-invite. Before creating one, ls .demos/ and reuse the existing name for the same flow instead of adding a near-duplicate.

Workflow

1. For form 1, read the views/routes of the flow to pick stable locators, and work out how to sign in (and with which test account) from the project's code and docs (AGENTS.md, README; /demo init sets that up). Then write the demo script (template below).
2. Record: $SKILL_DIR/scripts/record <name>. The first run builds the Docker image (tag hashed from the assets, so edits rebuild automatically). Before launching the browser the recorder checks that app.url (unset, http://localhost:$PORT from the environment or .env, else 3000) answers; if failed.json says app not reachable at <url>, start the app the way the project's AGENTS.md or README says, re-run, and stop what you started when done. If the app runs elsewhere, set app.url in the config.
3. Verify: extract 3-4 frames around the events and look at them. The t in events.json is seconds into the video.

   ffmpeg -loglevel error -y -ss 6.5 -i .demos/<name>/video.mp4 -frames:v 1 /tmp/frame.png

4. Report the video.mp4 and events.json paths.

Script

Runs inside the container next to lib.mjs, so always import ./lib.mjs.

Framing rules:

• Capture starts only when the demo script calls record(). Set the scene before it: browser open, logged in, starting page loaded. The video never shows the browser launching or an empty desktop.
• Never show the sign-in unless the user asks for it. Do it with plain Playwright before record(), not the animated helpers. Start on the page where the user's intent begins (e.g. /admin/users for "invite a user"), not on the home page.

Pacing: timings come from the pace config, so call helpers without timing arguments. Don't hover right before humanClick on the same target (the click already approaches it). Keep extra sleeps short: 300-500ms between steps, ~400ms after a new page loads.

import { launch, record, hover, humanClick, type, wheel, sleep, finish, APP_URL } from './lib.mjs'

const { page } = await launch()
await page.goto(`${APP_URL}/<sign-in-path>`, { waitUntil: 'networkidle' })
await page.getByLabel('<email-label>').fill('<test-account-email>')
await page.getByLabel('<password-label>').fill('<test-account-password>')
await Promise.all([
  page.waitForURL((url) => !url.pathname.startsWith('/<sign-in-path>')),
  page.getByRole('button', { name: '<submit-label>' }).click(),
])
await page.goto(`${APP_URL}/cart`, { waitUntil: 'networkidle' })
await record()

await hover(page, page.getByRole('link', { name: /Order summary/ }).last())
await humanClick(page, page.getByRole('link', { name: 'Checkout' }).first())
await wheel(page, 400)
await type(page, page.getByLabel('Coupon'), 'SAVE10')

await finish()

lib.mjs API: launch(), record({ lead }) (starts capture and the event clock; required, a demo script that never calls it fails), hover, humanClick, type (click then key by key), wheel(page, pixels), moveTo(x, y, ms) in screen coordinates, settle(page), mark(type, box) for custom telemetry, sleep, finish(ms).

Test accounts are project-specific: take them from the project's docs (AGENTS.md, README) or seed data, never guess.

Config

.demos/demo.config.mjs is optional and applies to every demo. Leave it out unless the user wants something different: the defaults are the house style. Keys are deep-merged over the defaults; unknown keys, wrong types, or out-of-range values fail the run before recording.

Create it with every default documented in place with /demo init ($SKILL_DIR/scripts/config), then edit or delete lines. It never overwrites an existing file. The file's comments are the reference for each key; defaults and pace presets live in $SKILL_DIR/assets/config.mjs. The top-level keys are app.url (where the app is served; unset, http://localhost:$PORT from the environment or .env, else 3000; localhost is the host's localhost: shared host network on Linux, forwarded to host.docker.internal on macOS), pace (normal or slow, or single timing overrides), theme (system follows GNOME or macOS appearance, light, dark), browser.locale, browser.window (width and height as screen fractions, window centered), desktop.dockAutohide, timeouts, and video.

Gotchas

• Scope locators: button.btn-cta can match a header copy first (off-screen, negative y). Prefer getByRole with a name, and .first()/.last() deliberately.
• Measure right before acting. The helpers scroll into view, wait for scroll to settle, and re-measure before clicking; custom moveTo code must do the same.
• Scroll only with wheel(). Never xdotool click 4/5: an event escaping the window switches the xfwm4 workspace and blacks out the rest of the video.
• End with finish(), never browser.close(), or the last seconds show an empty desktop.
• Form submits that navigate: wrap the click in Promise.all with page.waitForURL(...), then wait for the result (toast, row) before finish().
• The cursor is the real X pointer, so <dialog> top layer and (hover: hover) CSS just work.
• Native Chrome bubbles (translate, save password, "Change your password" leak warning after login, autofill suggestions under inputs) are disabled via profile prefs in launch(). If a new one shows up in failed.png, add its pref there.

Failures

Any error (missing locator, timeout) fails fast (5s action timeout) and the run exits non-zero with no new video. Read .demos/<name>/failed.json (error, stack, page URL, t, events so far), look at failed.png (the display at failure time; not written when the config is invalid or the app is unreachable, since those fail before the display starts), fix the demo script, and re-run. Never hand over a video from a failed run.

events.json

{
  "video": "video.mp4",
  "fps": 30,
  "width": 1440,
  "height": 900,
  "events": [{ "t": 6.504, "type": "click", "rect": { "x": 12, "y": 346, "w": 195, "h": 36 } }]
}

type is hover, click, type, or scroll. rect is in screen (video) pixels, measured when the pointer starts moving toward the target, so t leads the visible action by about 0.7s. scroll has rect: null.