How I Set Up Hermes to Write a Daily Tech News Report

I wanted a tech news briefing that didn’t suck. No algorithm-curated clickbait, no “you won’t believe what happened next,” no recycled press releases. Something sharp, analytical, like a senior analyst briefing a colleague. And I wanted it to show up in my Discord automatically, every few hours, without me lifting a finger.

So I made Hermes do it.

Here’s how.

The Cron Job

The core setup is a single cron job in Hermes:

Name:        tech-news-insights
Schedule:    15 */4 * * *   (every 4 hours at :15 past)
Model:       deepseek-v4-flash
Toolsets:    web, terminal
Deliver to:  Discord (dedicated channel)
Skill:       news-summary

That’s it. One cronjob create command and Hermes handles the rest. Every 4 hours, the scheduler spins up a fresh session, loads the news-summary skill, and drops the agent into the prompt.

The Prompt

The prompt is the secret sauce. It’s not a vague “summarize tech news today” — it’s a structured four-step workflow with hard constraints:

Step 1: Establish the time window. First thing the agent does is run date -u and date -u -d '4 hours ago'. This is the filter boundary. Nothing older than 4 hours gets through.

Step 2: Multi-source search. Three parallel searches with different phrasings to catch stories that one wording might miss:

web_search("breaking tech news today")
web_search("tech news last hour site:techcrunch.com OR site:theverge.com OR site:arstechnica.com OR site:reuters.com/technology")
web_search("technology news just now")

The site: operators constrain results to outlets with proper editorial standards — no Medium blog posts or SEO spam.

Step 3: Time-filter extraction. For each story, the agent extracts the full article via web_extract and checks the publication timestamp. If it’s older than 4 hours or the date is ambiguous, it gets dropped. This is the hardest part to get right — search engine dates are unreliable, and news sites use inconsistent timezone formats.

Step 4: Summarize and analyze. Each qualifying story gets:

• A 2-3 sentence summary
• A 1-2 sentence analytical insight connecting it to broader trends, market dynamics, or geopolitics
• A source link

The report ends with a “horizon scan” — a single line predicting what to watch for next based on the current signals.

The Output

The result drops into my Discord every 4 hours under the brand 🔬 Tech Radar. Here’s what a real output looks like:

🔬 Tech Radar — Friday, May 08, 2026 at 07:07 GMT+7

---

**OpenAI launches new voice intelligence features in its API**
OpenAI shipped new voice intelligence capabilities in its developer API, expanding what builders can do with real-time voice interactions.

💡 This is OpenAI commoditizing a layer that startups like ElevenLabs and Hume have been building on top of. Expect a wave of consolidation in the speech AI space as the platform play wins over point solutions.

🔗 https://techcrunch.com/...

---

**Kodiak AI raises $100M at a steep discount, stock tumbles 37%**
Autonomous trucking company Kodiak AI raised $100 million in a down round...

💡 Down rounds in AV are contagious signaling. If a public autonomous trucking company with real deployments has to eat this much dilution, private AV startups raising right now are getting absolutely hammered on valuation.

---

**Horizon scan:** Watch the autonomous vehicle sector for spillover from Kodiak's down round. On the AI voice front, OpenAI's API play sets up a pricing war that will shake out most standalone voice-AI startups within 6 months.

The tone is what makes it work: sharp, slightly opinionated, grounded in data. No “delve,” no “landscape,” no “it’s worth noting.” I explicitly banned those in the prompt because every AI loves them and they make everything sound like a McKinsey deck.

Why 4 Hours?

Daily news is too stale. By the time you read about a SoftBank deal or an OpenAI launch in a morning newsletter, the market has already priced it in.

Four hours hits the sweet spot:

• Fresh enough to catch Asian market open, European morning, and US afternoon in separate cycles
• Long enough that the window isn’t empty (midnight UTC often has zero breaking tech news)
• Not spammy — 6 reports per day is manageable for a dedicated Discord channel

If a window has no qualifying news, the agent responds with exactly "No breaking tech news in the last 4 hours." and exits. No padding, no reaching for older stories, no hallucinated filler.

The Skill

The news-summary skill provides the agent with backup sources — BBC, Reuters, NPR, and Al Jazeera RSS feeds with pre-built curl commands for parsing. The agent uses these as supplementary data when web search returns thin results or needs cross-verification.

1
2
3
4
5

# BBC Technology
curl -s "https://feeds.bbci.co.uk/news/technology/rss.xml" | \
  grep -E "<title>|<description>" | \
  sed 's/<[^>]*>//g' | \
  head -30

RSS is deliberately old-school. No API keys, no auth, no rate limits, no provider lock-in. It’s the Unix philosophy applied to news aggregation: small, composable tools that do one thing well.

What I Learned

1. Prompt engineering for consistency is real. The first version of the prompt was vague (“find interesting tech stories”). The agent would return 12 articles spanning 72 hours, half of them about crypto scams. Adding hard time constraints and explicit output format rules was the difference between noise and signal.

2. The model matters for analysis quality. When this job ran on deepseek-v4-pro, the analytical insights were genuinely sharp — connecting SoftBank’s moves to semiconductor supply chains, reading between the lines of EU regulatory retreats. The switch to deepseek-v4-flash reduced cost but traded depth for speed. For a personal news brief, I’ll take the savings. For client-facing reports, I’d pay for the bigger model.

3. Delivery destination matters. The report goes to a dedicated Discord channel, not my DMs. This means I can mute it when I’m heads-down, scroll back through history to catch up, and search for keywords (“SoftBank,” “ASML,” “down round”) to find specific stories later. Separating signal channels from general chat is an underrated productivity technique.

4. The 4-hour filter is brutal but necessary. About 30% of runs return “no breaking news” because the window falls during a quiet period (2-6 AM UTC, weekends). The agent sometimes needs 2-3 search retries with different query phrasings to find a single qualifying story. Without the filter, it would happily regurgitate yesterday’s news forever.

The Stack

Next Steps

The setup works, but there’s room to grow:

• Source diversity: Add Ars Technica, The Register, and Hacker News for deeper coverage
• Sentiment tracking: Tag stories as bullish/bearish and track which narratives are gaining momentum over multiple cycles
• Telegram delivery: Discord is fine for me, but Telegram would make this portable
• Voice briefings: The skill already has TTS support built in — trigger an audio summary for the morning commute

For now, though, I get six sharp, time-filtered, analytically-grounded tech briefings per day delivered to my Discord, and I never had to configure a single RSS reader, newsletter subscription, or news app.

That’s the promise of AI agents: not just answering questions, but doing the recurring cognitive work so you don’t have to.