social-signal-sweep

name	social-signal-sweep
context	fork
description	Recency-bounded sweep of what people publicly say about a topic over the last N days (default 30) across GitHub, Hacker News, Reddit, and Stack Overflow, plus YouTube and X via browser. Produces a lead sheet (platform, date, signal, link, what-to-verify) to seed deeper research. Strictly leads-only: social chatter is never citable evidence and must be verified before use. No engine, keys, or cookies; uses web fetch, the GitHub tools, and the simple browser. USE FOR: last 30 days, last N days, what are people saying about X, recent discussion, recency sweep, social signal, community sentiment, Reddit search, Hacker News search, GitHub activity last month, Stack Overflow recent questions, lead generation for research, find leads. DO NOT USE FOR: citable evidence or facts (use research-analyst VERIFY + citation-integrity), verifying a claim (use citation-integrity), authenticated or private scraping (use authenticated-web-extraction), exporting mail or calendar (use the outlook-* skills).

Social Signal Sweep

Survey what people are publicly saying about a topic over a bounded recent window (default 30 days) across community platforms, and return a lead sheet that a deeper investigation can act on. The output is a list of places to look, not a set of facts. Pair it with the research-analyst agent: this skill feeds its SOURCE phase; the agent's VERIFY phase decides what, if anything, is true.

Prime directive: leads, not evidence

Everything this skill returns is tier-8 (social / forums / chatbots) — a lead, never citable evidence. A post with 5,000 upvotes is still one anonymous assertion. Engagement measures attention, not accuracy. Do not state a sweep result as fact, do not let it raise a claim above Weak confidence on its own, and do not paste a social link as a citation. Every lead carries a "what to verify" so the next step is always to trace it to a primary source.

This is the opposite contract from a consumer "what's trending" tool. The value here is discovery — surfacing angles, contested points, version numbers, and named primary sources you would otherwise miss — handed to a discipline that treats all of it as unproven until verified.

When to use

• Scoping a new investigation and you need to know what angles, complaints, and claims are circulating before committing to a research question.
• Tracking reaction to a recent release, incident, CVE, paper, or announcement.
• Finding named primary sources (a maintainer's commit, a vendor advisory, a paper) that social posts point at — then leaving the social post behind.

When not to use

• You need a fact you can cite → go straight to research-analyst VERIFY and citation-integrity.
• You need to verify one specific claim → use citation-integrity.
• The data is behind a login → use authenticated-web-extraction.

Inputs

Input	Default	Notes
Topic	required	Exact entity/string. Disambiguate collisions (see Edge cases).
Window N days	30	Any integer. --days=7 weekly, --days=90 quarterly.
Platforms	GitHub, HN, Reddit, Stack Overflow	Add the browser tier on demand.

Workflow

1. Pre-flight the topic. Run the keyword-trap check (Edge cases) before any fetch. A doomed query wastes a sweep; reframing costs one sentence.
2. Compute the cutoff. Derive the epoch and ISO date for N days ago (see Time-bounding).
3. Sweep each platform with its recipe below. Prefer web/fetch against the public JSON endpoints; fall back to openSimpleBrowser only when blocked.
4. Dedup and cluster. Collapse cross-posts and the same story told on three platforms into one lead; a multi-platform cluster is a stronger lead than any single post.
5. Emit the lead sheet (format below). Every row gets a "what to verify".
6. Hand off. Route the sheet into research-analyst SOURCE → VERIFY, or hand individual citable-looking leads to citation-integrity.

Platform recipes

Public access patterns drift (Reddit lockdowns, X API changes). Treat each endpoint as the current no-auth path and fall back to the browser when a fetch returns 403/429. Never fabricate results to fill a gap — a blocked platform is a documented gap, not an empty row.

GitHub — highest reliability (native tools)

The only tier where the existence of an artefact (commit, release, advisory) is itself near-primary; the discussion around it is still a lead.

• Issues/PRs in the window: search with created:>=YYYY-MM-DD or updated:>=YYYY-MM-DD via the GitHub search tools (githubTextSearch).
• Releases and release notes: read them with the githubRepo tool.
• Repo trajectory: recent activity via pushed:>=YYYY-MM-DD; star movement as a coarse attention signal.
• Use github for issue/PR/discussion bodies and comments.

Hacker News — Algolia API (free, no auth)

Best date-bounding of any platform. Fetch JSON:

https://hn.algolia.com/api/v1/search_by_date?query=<topic>&tags=story&numericFilters=created_at_i>=<epoch>

Swap tags=comment for discussion. Each hit gives points, num_comments, created_at_i, and objectID → https://news.ycombinator.com/item?id=<objectID>.

Reddit — public JSON (rate-limited)

https://www.reddit.com/search.json?q=<topic>&sort=new&t=month&limit=50
https://www.reddit.com/r/<subreddit>/search.json?q=<topic>&restrict_sr=1&sort=top&t=month

Returns created_utc, score, num_comments, permalink. Reddit frequently returns 403/429 to non-browser requests. Set a descriptive User-Agent if the fetch tool allows; otherwise fall back to openSimpleBrowser on the equivalent old.reddit.com/search URL and scan visually. Expand to category-peer subreddits, not just the brand subreddit, or you miss cross-product discussion.

Stack Overflow / Stack Exchange — public API (throttled)

https://api.stackexchange.com/2.3/search/advanced?fromdate=<epoch>&order=desc&sort=creation&q=<topic>&site=stackoverflow&filter=withbody

Returns score, answer_count, creation_date, link. Switch site= for other Stack Exchange networks (serverfault, superuser, dba).

Browser-only tier — lower confidence (YouTube, X)

No stable no-auth JSON. Use openSimpleBrowser, scan manually, mark every lead "manual confirmation only":

• YouTube, sorted by upload date: https://www.youtube.com/results?search_query=<topic>&sp=CAI%253D
• X/Twitter: no reliable public no-auth API. Prefer a general web-search fetch for reactions; do not depend on third-party mirrors. This is the weakest tier — keep it optional.

Output: the lead sheet

# Social Signal Sweep: <topic> — last <N> days (as of <UTC date>)

**Scope:** <platforms swept> · **Cutoff:** <ISO date> · **Tier:** 8 (leads only)

| # | Platform | Date (UTC) | Signal | Gist | Link | What to verify |
|---|---|---|---|---|---|---|
| 1 | HN | 2026-05-30 | 412 pts, 233 cmt | Claims v17 broke replication | <url> | Reproduce against v17 release notes |
| 2 | GitHub | 2026-06-02 | issue, 18 👍 | Maintainer confirms regression | <url> | Read the linked commit/PR diff |
| 3 | Reddit | 2026-06-01 | 1.2k up | "Everyone is switching to X" | <url> | Find named adopters; treat as anecdote |

## Gaps and caveats

- <platform> returned <403/429/zero> — not swept / thin coverage.
- Topic disambiguation applied: <note>.

## Suggested next steps

- Tier-1 leads to verify first: <#, #>.
- Hand to `research-analyst` SOURCE; route citable-looking leads to `citation-integrity`.

Keep "Signal" as the platform-native engagement number (upvotes, points, comments, 👍). Never present it as evidence weight.

Time-bounding

Compute the cutoff once, reuse for every platform. HN and Stack Exchange want Unix epoch seconds; GitHub wants an ISO date:

$cutoff = [DateTimeOffset]::UtcNow.AddDays(-30)
$cutoff.ToUnixTimeSeconds()        # epoch for HN / Stack Exchange numericFilters / fromdate
$cutoff.ToString('yyyy-MM-dd')     # ISO date for GitHub created:/updated:/pushed: qualifiers

For Reddit, prefer the t= bucket (hour/day/week/month/year) and then filter the returned created_utc against $cutoff for sub-bucket precision.

Edge cases

• Keyword-trap topics (reframe before sweeping):
  • Demographic-shopping phrasing ("gift for a 42-year-old man") — real posts use relationship + hobby + budget, not the literal phrase. Ask for those first.
  • Bare numbers that collide ("42", "the 100") — strip the number unless it is load-bearing (GPT-4 keep; 40-year-old drop).
  • Tutorial phrasing ("how to use Docker") — people post "my Docker setup", not the tutorial title. Reframe to discussion vocabulary.
  • Generic single nouns ("sneakers", "coffee") — ask for the specific facet.
• Ambiguous entity (same name, different thing) — add a qualifying term and note the disambiguation in the sheet's caveats.
• Blocked platform (403/429) — fall back to the browser, else record the gap. Do not invent rows.
• Zero results — report it honestly; widen the window or add synonyms. A thin sweep is itself a finding, not a failure.

Relationship to other building blocks

• research-analyst — this skill is its tier-8 lead generator in the SOURCE phase. Leads stay Weak/Speculation until VERIFY triangulates them against higher-tier sources.
• citation-integrity — when a lead points at a concrete source (a paper, advisory, commit), hand that source to citation-integrity before any claim built on it is graded Established/Probable.
• authenticated-web-extraction — use it instead whenever a platform requires login; this skill is public-data-only by design.