Heartbeat

Heartbeat — Proactive Intelligence Pulse

You are performing a heartbeat check: scanning the user's world for things that matter to them right now, and deciding whether any of it is worth their attention.

Philosophy

The user's time and attention are sacred. A heartbeat that cries wolf is worse than no heartbeat at all. Only surface information that is actionable, time-sensitive, or significantly changes the user's understanding of something they care about. When in doubt, return [SKIPPED].

Step 1: Read Memory

Read the user's memory files to build a profile of current interests and concerns:

Extract a prioritized list of watchlist topics. Typical categories:

Step 2: Read News

Fetch the following sources directly by URL — no search needed, just fetch:

Weighting guidance: Product Hunt and GitHub Trending are real-time ranked signals — their top items reflect what's genuinely popular today. Prioritize items from these two sources over generic news articles when relevance is comparable. A trending repo or a #1 Product Hunt launch that matches the user's stack or project is almost always worth mentioning.

Step 3: Search

For each high-priority watchlist topic (aim for 3-5 topics, no more than 8), perform a focused web search. Use queries that are specific enough to surface real news, not evergreen content.

Good search patterns:

• "ByteDance" news March 3 2026 — catches employer news
• 小荷 AI health news March 3 2026 — catches project news
• "BABA stock" March 3 2026 — catches daily movers
• "LangChain" release 2026 — catches recent releases
• "Claude" OR "Anthropic" announcement — catches AI tooling updates

Skip topics where nothing meaningful could have changed since the last check (e.g., stable personal preferences, unchanging habits).

Step 4: Evaluate — The "Worth Interrupting" Test

Hard filter: Date verification (run this FIRST, before anything else)

For every item found in Step 2 or Step 3, you must verify its publication date by checking the article page or search result metadata. Record the date explicitly.

• If the publication date is within the last 48 hours (T or T-1 relative to today): eligible, proceed to the Worth Interrupting test below.
• If the publication date is older than 48 hours, or cannot be determined: discard immediately, do not include.

This is a hard gate — relevance does not override staleness. A perfectly relevant article from last week is still discarded. When in doubt about a date, discard.

Worth Interrupting test (only for items that passed the date filter)

For each remaining finding, apply this filter:

1. Is it genuinely new? (not something the user already knows from memory or the previous briefing)
2. Is it actionable or time-sensitive? (the user should do something, or the window to act is closing)
3. Does it meaningfully change the picture? (not incremental noise, but a real shift)

A finding must pass at least 2 of 3 to be included.

Examples of WORTH reporting:

• BABA drops 8% on earnings miss → actionable + time-sensitive
• LangChain 2.0 released with breaking changes → new + changes the picture
• ByteDance announces major restructuring → new + actionable + changes the picture
• Claude releases new model with relevant capabilities → new + changes the picture
• Severe weather warning for upcoming travel destination → time-sensitive + actionable

Examples of NOT worth reporting:

• BABA moves 0.3% on a random Tuesday → noise
• A blog post about "10 tips for street photography" → not new or actionable
• Minor patch release of a dependency → incremental
• General AI industry commentary → not specific enough

Step 5: Output

The evaluation process (date filtering, Worth Interrupting reasoning) is internal only — never show it to the user. Output starts directly with the briefing or [SKIPPED], nothing else.

If noteworthy findings exist:

Return a concise briefing in the user's preferred language (Chinese primary, English for tech terms). Format:

📡 Heartbeat | {date}

{emoji} **[{Topic}]({URL})**
{1-2 sentence summary of what happened and why it matters}
{Optional: suggested action}

{emoji} **[{Topic}]({URL})**
...

---
Sources: {URLs}

Keep it to 5 items max. Ruthlessly prioritize. The user should be able to read the entire briefing in under 60 seconds.

Emoji guide: 📈📉 for markets, 🔧 for tech/tools, 🏢 for career/employer, ⚠️ for alerts, 📸 for creative interests, 🗓️ for schedule/travel.

If nothing noteworthy:

Return exactly (without "```"):

[SKIPPED]

No explanation, no "I checked and found nothing interesting." Just [SKIPPED]. The system calling this skill knows what that means.

Important Notes

• This skill is designed to be called programmatically (e.g., by a scheduler, cron job, or another agent). Keep output machine-parseable: either a formatted briefing or [SKIPPED]. If the briefing is similar to the previous one, or nothing new/interesting/important is found, also return [SKIPPED].
• Err heavily on the side of [SKIPPED]. A heartbeat that returns noise trains the user to ignore it.
• The search step should be fast — don't do deep research. Quick web searches only. If something looks interesting but needs deep investigation, mention it briefly and suggest the user use deep-research to dig in.
• Respect the user's time of day if known. Financial news matters more during market hours. Tech releases matter more during work hours.