Trending Content OS
Trending Content OS is a health and wellness editorial radar that combines Google News, Google Trends, safety routing, and content opportunity scoring into a dashboard/report interface.
Local Setup
cd /Users/meaganmorris/trending-content-os
npm install
cp .env.example .env
npm run dev
Open:
http://localhost:3000
Pipeline Commands
npm run prefetch
npm run pipeline
The web app includes controls for starting a new report and refreshing the radar.
Locally, those controls run Python directly. In Vercel, they trigger the GitHub Actions workflow in .github/workflows/run-pipeline.yml.
In production, mutating run/scan endpoints require RUN_CONTROL_TOKEN. Enter that same value in
the dashboard's "Run token" field before starting a run. See VERCEL_DEPLOY.md for deployment
notes.
Quality Checks
Before pushing application changes, run:
npm ci
npm audit --audit-level=high
npm run build
python -m py_compile run_pipeline.py google_trend_radar.py run_tracked_topics.py
The same checks run in GitHub Actions via .github/workflows/quality-gate.yml.
Reading a run
The dashboard leads with the funnel — how many signals were reviewed, how many survived, how many
were cut — because the ratio (typically ~144 reviewed to ~9 retained) is the point of the system.
Under it:
- Trend strip — reviewed vs. retained for the last 14 runs.
- Run Insight — the model's own notes about the run.
- Priority Board — click any topic to expand its angle, working headline, why-now, and next
steps, with a "Copy brief" button. Filter by priority or search across angles and headlines. A
"day N" chip marks a story that has carried across consecutive runs.
- Rejected topics — collapsed by default, grouped by rejection reason.
- Why topics get cut — the same rejections aggregated across every run, with each reason's
share, whether it is rising or falling, and which config it implicates. Currently ~53% of all
rejections are off-category, which points at
GOOGLE_NEWS_QUERIES rather than at the filter.
- Generated runs — click any past date to load that whole run into the page
(
/?run=YYYY-MM-DD); the standalone HTML file is still there as a downloadable snapshot.
Each retained topic is also fact-checked against the evidence the run collected, and carries a
verdict badge — verified, minor issues, check before writing, or unverified (meaning the run's own
evidence did not cover the claim, which is normal for search-interest topics). Pass
--skip-fact-check to skip that pass.
Expanded rows also carry the source links, any ⚠️ integrity notes, and a one-line reason behind each
score (hover the score to see it). The run header shows what the run cost to produce.
Each run's numbers come from outputs/daily_newsroom_dashboard/extraction_<date>.json, the parsed
run data. raw_extraction_<date>.txt beside it is the raw model text kept for debugging — it is
fence-wrapped and was never valid JSON despite previously carrying a .json extension.
Headlines in the dashboard's "Google News Radar" panel link out to the original article, using
the Link: line the pipeline writes under each headline in
outputs/daily_newsroom_dashboard/serp_signals_*.md. Signal files generated before that line
existed have no URLs, so their headlines stay plain text until the next run.
Ad-hoc Trend Scanner
The dashboard's "Trend scanner" panel (#scan) runs google_trend_radar.py on demand for any
topic — not just health/wellness, and with no Anthropic API cost (pure SerpAPI + code-side trend
scoring). It's a separate tool from the daily automated pipeline above; see the "Known limitation"
/ Engineering Notes sections in CLAUDE.md for how it differs from run_pipeline.py.
The beauty profile additionally cross-checks Reddit (free, no API key) as a second signal
source, tags candidates with ingredient/benefit/concern labels from configs/beauty_taxonomy.yaml,
and computes an opportunity_score estimating whitespace (high trend interest, low existing
coverage) alongside the base radar_score — see CLAUDE.md's Engineering Notes for details.
There are also focused, single-topic profiles for each of the categories the daily pipeline
tracks depth for: nutrition, fitness, food-safety, diet, weight-loss,
mental-health, gut-health, womens-health (perimenopause, menopause symptoms, hormone
therapy, PCOS, endometriosis, period health, fertility; for women-focused publications; uses extra relevance anchors like PMDD, endometrial, estrogen
so adjacent conditions aren't dropped as off-topic) (plus the existing wellness/health/ai/beauty) — pick one
from the Trend Scanner's profile dropdown, or --profile nutrition on the CLI, to scan just
that category instead of the broad wellness/health blend.
Second batch of health profiles, each with its own weekly run-<profile>-radar.yml (Thursday-Saturday,
13:00-13:30 UTC):
| Profile | Seeds |
|---|
sleep | sleep apnea, insomnia, sleep tracking |
heart-metabolic | blood pressure, cholesterol, blood sugar, prediabetes, longevity |
brain-aging | brain fog, cognitive decline, dementia, healthy aging, bone density, sarcopenia |
adhd-focus | adult ADHD, ADHD medication, focus (extra anchors for drug names like Adderall/Vyvanse, AuDHD, executive function) |
supplements | creatine, collagen, electrolytes, vitamin D, iron deficiency, omega-3 |
chronic-conditions | long COVID, autoimmune, thyroid, migraine, chronic fatigue, inflammation |
public-health | measles, flu season, RSV, bird flu, FDA approval, drug recall, vaccine schedule |
habits-body | cold plunge, red light therapy, step count |
screening-prevention | colonoscopy, mammogram, skin cancer screening |
The most recent scan's plain keyword list (topic, stage, radar/opportunity scores, tags) also
shows up inline on the main dashboard under "Trending keywords" (#keywords), so you don't have
to open a separate HTML file to see what's trending.
A separate scheduled job, .github/workflows/run-beauty-radar.yml, runs the beauty profile
automatically once a week (Monday mornings) and commits the results, so "Trending keywords"
stays current without anyone clicking "Scan" — edit the cron line in that file to change how
often it runs.
Each of the other 17 profiles has its own weekly scheduled workflow too
(run-nutrition-radar.yml, run-fitness-radar.yml, run-food-safety-radar.yml,
run-diet-radar.yml, run-weight-loss-radar.yml, run-mental-health-radar.yml,
run-gut-health-radar.yml, run-womens-health-radar.yml, plus one per profile below), staggered across Monday-Saturday mornings so they don't all run
(and push) at the same time. All of them show up individually in the GitHub Actions tab and
can be triggered manually from there ("Run workflow") in addition to their schedule.
Locally the Scan button runs Python directly. Deployed, it dispatches
.github/workflows/run-radar.yml — a generic, on-demand counterpart to the per-profile workflows
that accepts any topic/profile/geo — which commits its results back and triggers a fresh deploy.
Either way you need a SERPAPI_API_KEY: locally in .env, deployed as a repo secret.
Vercel
This repo is Vercel-ready as a Next.js app. In v1, the web interface reads generated artifacts from:
outputs/daily_newsroom_dashboard/
The production button triggers GitHub Actions, which runs the Python pipeline and commits generated artifacts back to the repository.
See VERCEL_DEPLOY.md for deployment notes.
Trend Desk
The radars and the daily pipeline feed Trend Desk (trend_desk/), which decides what is worth
work and builds it: freelance pitch handoffs, Female Model outlines, Friday roundup items and SEO
briefs, approved by Meagan over Telegram. See trend_desk/README.md.
