The AI field moves too fast for any human to track comprehensively. I used to spend 30-45 minutes each morning scanning Hacker News, ArXiv, and various newsletters — and still missed things. So I built an agent that does it for me in under a minute.
The Pipeline
My agent runs daily on a cron schedule with three stages:
- Scrape & Collect — Gather raw content from 10+ sources (ArXiv, HN, RSS feeds, GitHub trends).
- ChatGPT — First Pass Filter — Process items in parallel, tag by category, and score relevance (1-10). Fast and cheap.
- Claude — Deep Analysis — Take the top ~15 items, produce summaries, cross-reference, and flag hype.
Stage 2: The Fast Filter
I use GPT-4o-mini for filtering — it's fast and cheap. The prompt returns structured JSON:
PROMPT = """
For each item, return:
- category: [research, product, infra, policy, opinion]
- relevance_score: 1-10
- one_sentence_summary: max 15 words
Items: {items_json}
Respond as JSON only.
"""From ~150 raw items, this filters down to 15-25 that score 6+. Cost: ~$0.05/day.
Stage 3: Claude's Deep Analysis
Claude Sonnet 4 receives the filtered items and produces a structured briefing with sections for major developments, notable research, infrastructure changes, and a "signal vs noise" assessment. Its strength is synthesizing information across sources and applying contextual judgment — often noticing connections ChatGPT misses.
Delivery & Cost
The briefing arrives via email at 7 AM (3-5 minutes to read), with an optional Telegram bot. Cost breakdown:
| Stage | Model | Cost/day |
|---|---|---|
| Scrape | — | $0.00 |
| Filter | GPT-4o-mini | ~$0.05 |
| Analysis | Claude Sonnet 4 | ~$0.15 |
Total: ~$6/month. Less than a cup of coffee for a personalized briefing better than any newsletter.
Iterative Improvements
- Feedback loop: Thumbs-up/down on items adjusts scoring weights weekly.
- Deduplication: Hash-based dedup avoids processing the same story from multiple sources.
- Historical context: Claude has access to a rolling 30-day index for tracking evolving stories.
- Weekend mode: Synthesizes the past 7 days into a longer, reflective piece.
The lesson: AI agents don't need to be complex. The best ones solve a specific, recurring pain point with a simple pipeline. Start small, measure value, and iterate.
Interested in applying AI agents to your business? I help companies build custom autonomous pipelines — from research to production deployment. Reach out to discuss your project →