ahastudio/til/geeknews-reactions

Find the GeekNews discussion for a TIL document's source article and enrich the document with key GN-native community reactions, each linked to the exact comment.

This skill must actually run to completion every time it is invoked — NO EXCEPTIONS. GeekNews requires a browser (no public search API), which makes it the easiest of the three reaction skills to skip under time pressure or when a queue is long. Skipping it, or declaring "probably no thread" without searching, is a defect. Every invocation must end in one of two concrete states, ever:

• A GN thread was found and searched for native comments (woven in, or

confirmed none worth weaving), with the GN 토론: line added.

• No matching GN thread exists after an actual search with narrower/broader

keyword variants — report this explicitly ("GN 스레드 없음 확인") rather than silently omitting any mention of GN.

Never substitute "I already checked HN/Lobste.rs" for actually searching GN — each platform is searched independently, because GN's own userbase and discussion patterns differ from HN's and Lobste.rs's, and a thread on one platform says nothing about whether one exists on another.

Contents

Section titles match the headings below; search for a title to jump to it.

• Usage: /geeknews-reactions [file-path]
• 1. Identify the source URL and any existing GN link
• 2. Find the GN discussion: search the GN search box in a real browser (Strategy A); web search alone never proves absence; never sweep topic ids
• Verify the match: the source URL must appear in the topic page
• 3. Fetch the discussion page and comment anchors: topic pages are server-rendered, cid anchors
• 4. Filter out re-aggregated HN/Lobste.rs comments: skip the aggregator account
• 5. Select significant comments: native GN comments only
• 6. Add the GN link to the document header
• 7. Weave reactions into the document body: handle footnotes, one permalink per comment
• 8. Writing rules

Usage

/geeknews-reactions
/geeknews-reactions mac/some-tool.md

• Argument: path to an existing TIL file (optional; if omitted, use the file

currently open in the IDE, or ask the user)

Procedure

1. Identify the source URL and any existing GN link

Read the target TIL file. Extract the source URL from the 원문: line (or the first link at the top of the document).

If the document already has a GN 토론:, Show GN: or Ask GN: line, use that URL directly and skip to step 3.

2. Find the GN discussion

Two things about GeekNews, both verified:

• `/search?q=...` is JavaScript-rendered. curl and WebFetch both get

the page shell with zero results and HTTP 200. An empty result from it is not evidence of absence (see the web-fetching skill, «An Empty Result Is Not a Result»). The shell is also the same size for every query — roughly 52KB whether the term has 484 hits or none — so a length or diff heuristic will not reveal that the search never ran.

• Topic pages are NOT JavaScript-rendered. https://news.hada.io/topic?id=N

returns full server-rendered HTML including every comment and its cid anchor. No browser is needed for step 3.

So the browser is needed to find the topic id, and a web search engine is not a substitute for it: in one queue, domain-restricted WebSearch reported no GN thread for six documents whose threads existed, found afterwards through the GN search box. A negative conclusion from WebSearch alone is never allowed.

Strategy A — the GN search box in a real browser (primary)

Per the web-fetching skill, use a real browser (Claude in Chrome; fall back to agent-browser only when Chrome is not connected, and say so). Open https://news.hada.io/search?q=<keywords> and read the rendered page, not the HTML.

Search Korean keywords, because GN titles are Korean and often rewritten rather than translated. Try at least three phrasings per document: a literal rendering of the title, the product or project name alone, and a description of what it does.

To run several phrasings in one call, from a tab already on news.hada.io, load each query in a hidden same-origin iframe, wait about 2.5 seconds, and read d.body.innerText plus the topic?id= links:

const qs = ["<phrasing 1>", "<phrasing 2>", "<phrasing 3>"];
const res = [];
for (const q of qs) {
  const f = document.createElement("iframe");
  f.style.cssText = "position:fixed;left:-9999px;width:800px;height:600px";
  f.src = "/search?q=" + encodeURIComponent(q);
  document.body.appendChild(f);
  await new Promise(r => f.onload = r);
  await new Promise(r => setTimeout(r, 2500));
  const d = f.contentDocument;
  const count = (d.body.innerText.match(/검색결과 약 \d+개/) || ["no count line"])[0];
  const seen = new Set(), items = [];
  for (const a of d.querySelectorAll("a")) {
    const m = (a.getAttribute("href") || "").match(/topic\?id=(\d+)/);
    if (m && !seen.has(m[1])) { seen.add(m[1]); items.push(m[1] + " " + a.textContent.trim().slice(0, 40)); }
  }
  res.push(q + " [" + count + "] " + items.slice(0, 5).join(" ; "));
  f.remove();
}
res.join("\n").replace(/[?=&]/g, " ")

Two limits of that call: keep it to about six queries (the tool times out near 45 seconds), and strip ?, = and & from what it returns, because a return value that looks like query-string data is replaced wholesale with a blocked notice.

Wait before concluding zero. Immediately after loading, the count line is often absent because the results have not been injected yet, and that state is indistinguishable from a genuine zero. Before reporting absence, require one of these two:

• the 검색결과 약 N개 line with a count, and no matching topic among the

results, or

• an explicit no-result phrase in the body (결과가 없, 일치하는,

찾을 수 없).

If neither is present, the page has not finished rendering — read it again rather than writing a conclusion.

Run a control query. ripgrep returns about 500 results through this path. If the control returns nothing, the search is not running and no negative conclusion may be drawn from that session.

Strategy B — domain-restricted web search (supplement only)

WebSearch(query: "<한국어 주제 키워드>", allowed_domains: ["news.hada.io"]) can surface a topic quickly, but its index is incomplete and recent topics are often missing. Use it to add candidates, never to rule a thread out.

Verify the match

Fetch the topic page (step 3) and confirm the source URL appears in it. A matching title is not enough — GN carries many topics on the same subject from different sources.

If no candidate's source URL matches after Strategy A with three phrasings and a passing control query, report that GN has no thread and stop — do not guess a topic id.

Never sweep a range of topic ids. Fetching topic pages sequentially to hunt for one triggers an IP-level block: GN starts returning HTTP 403 with a 10-byte body for every request, across every language subdomain (es., ja., …), and it persists for hours. This has already cost a session the ability to read GN at all. If you are tempted to scan ids, the GN search box (Strategy A) is the answer instead.

3. Fetch the discussion page and comment anchors

Plain HTTP with a browser User-Agent is enough:

UA="Mozilla/5.0 (Macintosh; Intel Mac OS X 10_15_7) AppleWebKit/537.36"
curl -s -A "$UA" "https://news.hada.io/topic?id=<topic-id>" | python3 -c "
import re,sys,html
h=sys.stdin.read()
print('출처 URL 포함:', '<source-domain>' in h)
parts=re.split(r\"id=['\\\"]?(cid\d+)['\\\"]?\", h)
for i in range(1, len(parts), 2):
    t=re.sub(r'<script.*?</script>','',parts[i+1],flags=re.S)
    t=re.sub('<[^>]+>',' ', t)
    t=re.sub(r'\s+',' ', html.unescape(t)).strip()
    t=re.sub(r'^data-comment[^>]*>\s*','',t)
    m=re.search(r'▲\s*(\S+)', t)
    print(parts[i], '|', m.group(1) if m else '?', '::', t[:600])
"

Each comment element's id attribute (e.g. cid56704) is the exact anchor to use in footnote URLs: https://news.hada.io/topic?id=<topic-id>#<cid>.

If this returns HTTP 403 with a tiny body, you are IP-blocked; WebFetch on the same topic URL still returns the comment text and handles, but it strips HTML attributes, so you will get the comments without their cid anchors. In that state read the topic page inside the browser instead: from a tab on news.hada.io, fetch("/topic?id=<topic-id>"), parse it with DOMParser, and take every [id^=cid] element's id and innerText. The browser is not subject to the block the plain HTTP client hit, and the cid anchors are intact. Only if that also fails, read the discussion with WebFetch and say that it cannot be footnoted rather than guessing anchors.

4. Filter out re-aggregated HN/Lobste.rs comments

GN topic pages often embed a "Hacker News 의견들" or "Lobste.rs 의견들" block that just republishes comments from those other sites (attributed to the GN⁺ account, not a real GN user). These are NOT native GN reactions.

• Skip any comment block introduced by "Hacker News 의견들" or

"Lobste.rs 의견들" headings.

• Only consider comments posted by real GN usernames (not GN⁺) as

candidates for step 5.

• If the document already has /hackernews-reactions or

/lobsters-reactions footnotes covering the same underlying quote, treat it as already covered and skip it — do not duplicate the same reaction under a new GN footnote just because GN also republished it.

5. Select significant comments

Identify native GN comments that add something not already in the TIL document. Criteria:

• A concrete counter-argument, counter-example, or rebuttal
• A real-world anecdote, case study, or hands-on usage report (GN threads

often include Korean users' first-hand trial results)

• A useful distinction or re-framing of the article's concepts
• A notable criticism of the article's logic or scope
• A well-received alternative interpretation

Skip: pure praise, off-topic jokes, one-liners with no substance, or content already covered in the document (including content already covered via /hackernews-reactions or /lobsters-reactions).

Aim for 3–8 comments — GN threads are typically much smaller than HN ones. Quality over quantity; it is fine to integrate fewer than 5 if that is all the thread offers.

6. Add the GN link to the document header

If the document already has a Show GN: or Ask GN: line, that line is the GN discussion line; do not add a GN 토론: line next to it. Otherwise, if the document does not already have a GN 토론: line, add it after any existing HN 토론: / Lobste.rs 토론: lines, otherwise immediately after the 원문: line:

GN 토론: <https://news.hada.io/topic?id=TOPIC_ID>

7. Weave reactions into the document body

For each selected comment, integrate it into the most relevant existing section (## 분석, ## 비평, or ## 인사이트). Rules: