cd /news/artificial-intelligence/watching-hacker-news-turning-top-thr… · home › topics › artificial-intelligence › article
[ARTICLE · art-174147] src=lusob.github.io ↗ pub= topic=artificial-intelligence verified=true sentiment=↑ positive

Watching Hacker News: turning top threads into 6-minute short films

Developer @lusob built HN DOCS, a pipeline at hndocs.org that turns the most-voted Hacker News story of the last 24 hours into a six- or seven-minute narrated documentary film in English and Spanish, using Claude for research and scripting, ElevenLabs for voice, Remotion for 1080p rendering, and Cloudflare Pages and R2 for hosting at roughly 250 MB per language per film. The pipeline produces about 20 scenes and 85 to 95 shots per film, with each claim in Claude's research dossier labeled solid, disputed or debunked so the script only states solid claims as fact. The author reports that generating footage is easy but matching each shot to the narrated word is hard, and that filtering stock agencies out of the search and showing real Hacker News thread screenshots improved credibility.

recommendation

Have the agent write a dossier labeling each claim solid, disputed or debunked before scripting, and build a review page that lists mismatched shots with candidates instead of relying on one full-speed watch.

more in playbook: Designing agent systems
read3 min views10 publishedNov 22, 2026

A short documentary for each of the most popular stories on Hacker News. hndocs.org · code A Hacker News thread is where the interesting argument about a story happens. But the context you need to follow that argument (the history, the numbers, who said what) takes an afternoon of open tabs. So I built a pipeline that spends that afternoon for you and turns it into a six or seven minute film: narrated, with real footage, with the thread’s disagreements in it.

Each film is narrated in English and in Spanish over one single cut. There are no accounts; your list, ratings and progress stay in your browser.

How a film gets made #

  1. Pick the story. The most voted one of the last 24 hours, or one a reader suggested.
  2. Research. Claude reads the article and the Hacker News thread and searches the web. It writes a dossier where every claim is labeled solid, disputed or debunked, and the script is only allowed to state the solid ones as fact.
  3. Script and shot plan. About 20 scenes, and roughly 85 to 95 shots per film. Each shot starts on the word of the narration that names it.
  4. Voice. ElevenLabs. The Spanish version is condensed where it would overrun the English picture, and voiced again.
  5. Footage. Photos, video, screenshots of the actual thread when the narration mentions it, screenshots of the headline when it names an outlet, and animations designed for each film. When nothing real exists, an AI-generated image, marked as such in the credits.
  6. Review. A local page plays the narration with the shots at their real timing. For each one I pick: ok, remove it, find another. The pipeline applies the decisions.
  7. Render. Remotion, in 1080p, with depth maps to give some photographs real parallax.

What I learned #

  • Generating is easy; matching is hard. A shot that does not show what the narration says right then is worse than no shot. When one does not fit, the first thing to try is removing it and letting its neighbour run longer, not searching for a replacement.
  • Stock footage is a trap. Watermarks (even faint ones), people who are not in the story, generic keyboards, code from the wrong decade. I ended up filtering the stock agencies out of the search altogether and keeping a list of what counts as a bad shot.
  • Show the real thing. When the narration says “on Hacker News”, the screen shows that thread. It does more for credibility than any stock shot.
  • Labels beat vibes. Forcing every claim in the dossier to be marked solid, disputed or debunked changed what ended up in the scripts more than any prompt tweak did.
  • A review tool pays for itself. Watching a film once at full speed is not enough to find the ten shots that are wrong. A page that lists them with their candidates is.

What it runs on #

Claude Code did the research, the scripts and most of the code. Voices by ElevenLabs, video by Remotion, the site on Cloudflare Pages and R2 (a film is about 250 MB per language), visits counted with GoatCounter. Everything runs from a single Python CLI, pltv.

Caveats #

The scripts and the voices are AI, and the dossier checking is not infallible: if you find a mistake, tell me. The footage comes from many public sources and each film lists its credits; it is used to illustrate the story, and if something shown is yours and you want it credited or removed, write to contact@hndocs.org. HN DOCS is an independent project, not affiliated with Hacker News, Y Combinator, Anthropic or ElevenLabs.

Process #

I use Claude Code on the desktop. My own manual work was: creating the API keys (ElevenLabs, Pixabay, Cloudflare), getting the domain, and reviewing every film shot by shot on the review page. Same spirit as Andrej Karpathy’s HackerTube: describe what you want, and let the agent build the thing.

— @lusob

── more in #artificial-intelligence 4 stories · sorted by recency
── more on @hn docs 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
→ Live at https://your-agent.zahid.host ✓
Get free account → Pricing
from €0/mo · no card required
LIVE [news/watching-hacker-news…] indexed:0 read:3min 2026-11-22 · —