Emerging keyword monitor (newterm)
Reusable playbook for a bot that finds new searchable keywords / AI memes and sends Chinese broadcasts. No API keys belong in this skill — store them in the owner's .env or secret store only.
Schedule (Asia/Shanghai)
- Prescan (data only, no user message): ~08:20, 10:20, 16:20 → slots
morning/noon/evening - Broadcast: 09:30, 11:58, 17:50 (noon = new items only; silence if none)
- Night collect (no broadcast): ~00:30, 04:30
Prescan writes under prescan_<YYYYMMDD>_<slot>/. Broadcasts only summarize; they do not re-run the heavy dry scan.
Heat unit
Always report heat as GPTs × N = global past-7-day mean interest ÷ GPTs baseline (same idea as heat_ratio from keyword feeds that already provide it). Never dump raw Trends index numbers as the main figure.
Source rules (what enters the broadcast)
First-hand → broadcast when first seen; Trends is display-only (no gate):
- Social: TikTok / X / TikHub-style meme search
- Reddit (first appearance + human noise filter)
- Big-site inner pages from sitemap radar
- Own first-hand: tool directories / GitHub (high-velocity new repos) / GitHub Trending daily / Hugging Face
- CapCut public trending templates
- Product Hunt new products (Atom; optional GraphQL if owner has a token)
Still Trends-gated:
- Curated keyword feed with verdicts (e.g. 「值得做」「可以试试」) + heat_ratio ≥ ×0.2
- Root rising / Trends hot searches ≥ ×0.2
TikTok: ≥ ×0.05 formal push (must note approximate appear time inside 来源); below → 「TikTok 早期观察」, first sighting must still be listed.
Intent filters (not “no volume”)
Skip evergreen commerce/how-to, pure domain navigation, sports schedule noise, seasonal evergreen, brand shopping, one-day spikes that collapse, political news. For Reddit/CapCut, drop resolutions, quant names, pure param fragments; keep product / model / game / meme-like searchable terms.
Dedup
Track pushed spellings in pushed.json. Same-day re-push of the same spelling: no. Prior days still over gate: may re-push. Noon: only brand-new terms or new spellings of old topics.
Report shape (mobile-friendly, from 2026-10-25)
No tables. Never squeeze fields with /.
Lead (max 2 lines): title + one count line, e.g. 新增 23:新词 5 · 新写法 12 · TikTok早期 4. No batch IDs, gate essays, or prescan status walls.
Whitespace: blank line between sections; blank line before each numbered term.
Each formal term: bold **N. term** alone on a line, then Markdown list lines. Drop the redundant 新词: line (title already has the term).
- Required: 热度 (GPTs × N), 来源 (include roughly when it started rising), 是做什么的 (at most two short sentences: what it is; what searchers want)
- Optional one-per-line: 走势, 别称 (only if different from the title; replaces old 主流写法), 竞品网站, 推过记录, 哥飞, 风险
- Appear time is merged into 来源 (no separate 出现时间 line)
New spellings: shorter — 热度, 推过记录, 是做什么的; 来源 optional one line.
Empty sections: section title then 无 only.
Sections: 一新词;二新写法;三再推;四 TikTok+早期观察;五 Reddit/CapCut/Product Hunt/GitHub Trending;六大站内页;七过线但没推 (subheads by source, list lines; same-reason terms can share one line).
Owner-local sample file often lives at BROADCAST_FORMAT.md next to the pipeline.
Pipeline layout (owner rebuilds code locally)
Typical tree under a workspace dir (e.g. /workspace/newterm):
prescan.sh <slot>— orchestrates dry scan + TikTok + radarsnight_collect.sh— night-only collect- Collectors: sitemap radar, Reddit RSS, CapCut public
/template, Product Hunt Atom, GitHub Trending daily (radar_github_trending.py), TikTok merger (X + public tags + TikHub) - Trends client via a Google Trends explore/live proxy (prefer UTC day cut; drop incomplete last hour; empty→0). Fallback to unofficial Trends libs only if proxy fails.
- Keyword-batch API client for the curated feed (batches list + batch detail)
- State:
pushed.json, seen sets per radar, Trends cache
Cost controls (important)
CACHE_TTL_H≈18,ROOTS_TTL_H≈24so evening does not fully re-query morning- Noon/evening TikTok scan:
--no-trends(Trends once in morning) - Daytime sitemap:
--no-score(same as night) - Do not put SERP volume APIs on the daily cron
Env var names only (never pack values)
Owner supplies their own: Trends proxy key, TikHub (or equivalent), curated keyword API token if required, optional Product Hunt token, optional GH_TOKEN/GITHUB_TOKEN for GitHub API fallback. Scripts must read from env / .env and never echo secrets.
Sitemap big sites (example set; owner may edit)
Template/trend product sites + gaming content sites commonly watched for new inner pages: makeviral, getstarrd, viberot, trendswap, kavel, kie, opentrend, pokemongohub, esportsdriven, keengamer (plus snowball expansions). New inner page for an unpushed term → broadcast with no heat gate.
Services to connect (no secrets in template)
- Built-in or plugin X reads for recent meme posts (free public reads where available)
- Owner's Trends proxy / DataForSEO-style explore live
- Optional TikHub-class TikTok search API
- Optional Product Hunt GraphQL
- Curated SEO keyword batch API if used
- Optional GitHub token for Trending API fallback
When importing: ask the owner for which slots they want, which sources to enable, and where to put API keys — then create routines and a local .env without writing key values into memory or skills.