Tech Politics Brief

A proactive platform that surfaces worldwide tech-politics news from primary and secondary sources, pre-sorts it, scores it with a Claude layer, matches every item against your own theses, and publishes a daily Top 5 dashboard.

Topics it tracks: EU Chips Act & Chips Act 2.0 · EU AI regulation · digital sovereignty · AI industrial policy · US technology policy · semiconductor industry news.

How it works — a pipeline of named agents

fetch → cluster (novelty) → score → thesis-match → edit → render/notify

1. Collector (fetch.py) pulls ~20 feeds from config/sources.yaml (RSS + the keyless US Federal Register JSON API).
2. Novelty agent (cluster.py) groups the same story across sources and answers "is this already everywhere?" — more independent domains ⇒ less novel, but a primary-source break stays novel even when the trade press piles on.
3. Scoring agent (scoring.py) rates each story 1–5 on relevance, novelty, and carousel-worthiness.
4. Thesis agent (thesis.py) matches each item to your standing arguments in config/theses.yaml (supports / contradicts / extends / neutral).
5. Editor agent (editor.py) picks the daily Top 5 with topic diversity and writes a why-it-matters line and a carousel hook per pick.
6. Render / Notify produce a self-contained HTML dashboard, a dated markdown archive, and an email digest.

A parallel, independent pass — Topic Discovery (topics.py) — runs over the full cluster pool (not just the Top-5 candidates) to build output/topics.html: open-vocabulary topics (not the fixed keyword list above), split Europe vs. worldwide from article content, plus a Hidden & emerging section for themes trending above their own trailing average but not yet mainstream (needs ~1–2 weeks of daily runs to build a baseline — output/topic_history.json is the durable record, since it's what the daily workflow actually commits back to the repo).

No API key? It still runs. Every Claude-backed agent has a deterministic heuristic fallback, so the whole pipeline works offline. Set ANTHROPIC_API_KEY to switch scoring/editing to Claude (Haiku for per-item scoring, Sonnet for the editor) — the code path is identical.

Quick start

python3 -m venv .venv && source .venv/bin/activate
pip install -e ".[dev]"

technews run --once          # fetch, score, render → output/dashboard.html + output/topics.html
open output/dashboard.html   # (or xdg-open) view the Top 5
open output/topics.html      # what's being discussed, Europe vs. worldwide + emerging themes

technews show                # print the latest edition as markdown
technews fetch               # just fetch and report per-source counts
technews notify --dry-run    # render the email digest without sending

With Claude scoring:

export ANTHROPIC_API_KEY=sk-ant-...
technews run --once

Configuration (config/)

• sources.yaml — the source registry, 4 tiers (primary · institute · secondary · newsletter). Each source carries topics tags; tier feeds the scoring so primary sources anchor relevance and novelty. Add/replace feeds freely — unreachable ones are skipped, never fatal.
• theses.yaml — your arguments. Every item is matched against these; edit the claims and keywords to make the brief personal.
• settings.yaml — score weights, top_n, recency_days, cost cap (max_items_per_day), Claude model ids, and the Resend email from/to. Topic Discovery has its own cost guard (topic_scan_limit) and trend window (topic_history_days).

Email digest (Resend)

Email uses Resend: a single RESEND_API_KEY, good deliverability for automated mail, and self-delivery works without a custom domain. Set the key and recipient, then:

export RESEND_API_KEY=re_...
technews run --email

Without the key, --email / notify render the HTML but don't send.

Deploy (Vercel) + daily automation

The dashboard is a single self-contained HTML file — host it anywhere static. This repo ships a Vercel config (vercel.json, output dir output/).

.github/workflows/daily.yml runs the pipeline every morning (06:00 UTC), commits the dated markdown archive, emails the digest, and deploys the dashboard to Vercel. Add these repo secrets to enable the optional pieces:

Secret	Enables
ANTHROPIC_API_KEY	Claude scoring/editing (else heuristic)
RESEND_API_KEY	Email digest
VERCEL_TOKEN	Auto-deploy the dashboard

Cost

Hosting (Vercel), CI (GitHub Actions), and email (Resend free tier) are €0. The only variable cost is the Claude scoring layer — roughly €6–18/month at ~50–100 stories/day (Haiku for scoring, Sonnet for the editor, prompt-cached). The max_items_per_day cap and the Batch API keep it well under €20; the heuristic fallback is €0.

Tests

pytest -q          # offline: fixtures for fetch, clustering, scoring, thesis, render