# HackerTube-inspired comment-to-watchlist study


## Event and study boundary

Andrej Karpathy's [HackerTube post, 31 October 2026](https://karpathy.bearblog.dev/hacker-tube/)
describes using Claude tools to turn Hacker News comments into a video-discovery site. The author
reports roughly 40,000 recommending comments, 6,454 videos and 150 topics. Those are the author's
whole-site figures, not measurements from this sample. [Original HN discussion](https://news.ycombinator.com/item?id=50319510).
This package supports an independent study inspired by that idea. It does not evaluate HackerTube's
actual catalogue, reproduce its server or require the author's model. It uses ordinary public comments.

## Frozen sample and files

The query was `astronomy youtube.com` with `tags=comment`, `hitsPerPage=40` at the
[Algolia HN Search endpoint](https://hn.algolia.com/api/v1/search?query=astronomy+youtube.com&tags=comment&hitsPerPage=40). The first response was frozen on
2026-11-03T05:26:08.015008+00:00; its returned relevance order was retained. Within those 40 hits, every
comment containing at least one specific 11-character YouTube video ID in a `watch?v=` or `youtu.be/`
link was retained. Channel-only links were excluded by that mechanical rule. No selection was made
on whether a comment recommends, criticizes, or accurately describes a video. There are
**23 comments, 38 distinct comment–video pairs and 36 unique video IDs**.
The astronomy keyword is a sampling entry, not a label that every video is relevant to space science.
Some hits discuss other subjects or long lists. All exact video IDs in the selected original comments
were retained; topics, recommendation intent and suitability remain questions for analysis.

- `hn_comments.jsonl`: one record per selected comment. Official Firebase HTML is preserved
  without rewriting, with a plain-text convenience rendering, original links, root-story information
  and up to three parent items nearest first. A separately fetched root story does not fill any missing
  intermediate comments. `parent_chain_truncated` explicitly marks the limit. Search text is also
  preserved, so discrepancies with the subsequently fetched official comment remain visible.
- `comment_video_links.csv`: one row per distinct video ID mentioned in each original comment.
  A mention is not a recommendation label. Multiple comments may point to the same video.
- `video_metadata.csv`: one row per unique video ID. Titles and authors come only from successful
  public YouTube oEmbed responses. Unavailable or unchecked metadata remains blank with its status.

## Observed access and interpretation limits

Original source retrieval uses the [official HN Firebase API](https://github.com/HackerNews/API),
with comment links pointing back to `news.ycombinator.com`. All 23
selected comment texts were available when frozen; 3 parent chains hit
the three-level limit. Public oEmbed supplied metadata for 33 of
36 videos; the others retain their actual response status. Metadata success
does not establish playback, regional availability, accuracy or educational quality. No video was
downloaded or watched; full descriptions and transcripts are not provided. Metadata titles alone
may be insufficient to determine relevance, so uncertainty is a valid outcome.

This is a small, query-selected, non-blind sample. The same material can support screening and
later evidence review, but it has no separately prepared human gold standard. Do not call AI
self-check agreement human accuracy or estimate whole-site recommendation accuracy from it.
Duplicate IDs, unavailable metadata, ambiguous wording and absent context remain part of the sample.
Public comments belong to their original authors; preserve attribution and source links.

The data freeze records observed inputs and access states. It contains no pre-generated AI judgments,
desired error examples or promise that the final watchlist must have a particular size.
