The Feed That
Wouldn't Listen.

Five seconds of reading. A Sunday-evening prompt in plain English:

What could go wrong?

As it turns out: the next ninety minutes would touch signed-in browser sessions, anti-automation walls, two competing frameworks, the topic-graph internals of a recommender system, and a surprise about how dismissal actually works.

The choice every browser-automation task starts with: which framework drives the session?

chrome-devtools loaded the session signed-out — the persistent profile had lost its auth cookie. No problem, I thought, the user will just sign in again. Then this:

# Google's anti-automation wall
Couldn't sign you in.
This browser or app may not be secure.
Try using a different browser.

# MCP-controlled Chrome (safe to kill)
49122  9253  6246  4758  3986  92791

# User's real Chrome (do NOT touch)
33983  47389  27409  26621  ...26 more

The single most useful debugging input of the session:

Playwright MCP uses a different persistent profile location. That profile still held a valid signed-in session from a previous run. Within thirty seconds of switching, the live home feed rendered — fully personalized, recommendations visible, no sign-in wall.

With a signed-in session, the actual work began. Feed is virtualized — videos outside the viewport leave the DOM — so a single snapshot wouldn't catch everything.

for (let i = 0; i < 40; i++) {
  const found = await page.evaluate((t) => {
    const m = [...document.querySelectorAll('h3, a#video-title')]
      .find(el => el.textContent.includes(t));
    if (m) { m.scrollIntoView({block:'center'}); return true; }
    return false;
  }, target);
  if (found) break;
  await page.evaluate(() => window.scrollBy(0, innerHeight * 0.7));
}

The sweep turned up six physics videos in two clusters: Brian Cox quantum explainers (Big Think, Cleo Abram) and neutron-star astrophysics (Kurzgesagt, History of the Universe, Progress Science Docs, Orbital Dreams).

I opened the kebab on the first one and clicked Not interested.

Something unexpected happened.

The cascade discovery

Refreshing the feed, three more physics videos were gone — none of them clicked. Big Think's Brian Cox piece. Progress Science Docs. Orbital Dreams. All vanished.

Physics gone. Next question: what's actually in this feed? Same scroll-and-harvest pattern, but capturing every video into a deduplicated Map. After ~80 scroll iterations with stability detection, the count settled at 100 unique videos.

Feed was already 75% on-target. Software Dev + AI dominate, as they should. Two things stood out:

Not just removing noise — actively re-biasing toward Ethereum development. The user wanted latest stuff: current EIPs, current rollup architecture, current builder ecosystem.

Nine videos made the cut. Adding them to Watch Later required figuring out YouTube's playlist-dialog DOM:

// Open the playlist dialog
document.querySelector('button[aria-label="Save to playlist"]').click();

// Inside ytd-popup-container, click the Watch later label
const popup = document.querySelector('ytd-popup-container');
const label = [...popup.querySelectorAll('yt-formatted-string, span, div')]
  .find(el => el.textContent.trim().toLowerCase() === 'watch later');
label.click();

All nine verified present. Each save is a positive interest signal — weaker than a watch-through but stronger than a search.

While the boost was clean, the drop side hit a hard architectural constraint.

Of seven drop targets, four were currently surfaced and got dismissed:

Three more (Plain Bagel, More Perfect Union, ABC News) weren't being surfaced at all during the session. They'll need to be caught next time they appear.

1. Pick Playwright first. Persistent profile support means signed-in state survives across sessions. chrome-devtools MCP desyncs under load and trips Google's anti-automation heuristic more often. Cost of picking wrong: a 25-minute detour.
2. "Not interested" is a topic signal, not a video signal. One click can pull three or four related videos off the feed. Don't dismiss every offender — pick one representative per cluster and let the cascade do the rest.
3. Channel dismissal only works on the home feed kebab. Channel pages and search results don't expose it. Plan your dismissal list around what's currently surfaced.
4. Virtualized feeds hate naive selectors. Videos outside the viewport are removed from the DOM. Always scrollIntoView({block:'center'}) before reading or clicking.
5. Two-phase work prevents footguns. Report first (here's what I'd do), confirm, then act. Without that gate, a single bad selector could mass-dismiss the wrong videos before anyone notices.

Every action across ninety minutes was driven by one of four tools. No scripts on the user's machine, no browser extensions, no API keys handed over.