pulse

---
name: pulse
description: Use when the user wants multi-source recency intelligence on a topic — 'pulse on X', 'what are people saying about X', trend discovery, competitor research, or sentiment analysis.
category: Research
version: 1.0.0
tools: []
---

# Pulse — Multi-Source Recency Research

Synthesize what people are currently saying about a topic — across Reddit, Hacker News, and the open web via `web_search`/`knowledge_base_search` — within a configurable recent time window. Deliver a single coherent chat briefing with citations, engagement signals, and cross-source pattern analysis. This captures the **current conversation**, not the canonical reference.

**Trigger phrases:** "pulse on X," "what's happening with X," "what are people saying about X," "current conversation about X," "take the pulse of X," "trending: X," or requests for competitor research, trend discovery, tool comparisons, or audience-sentiment analysis with a recency flavor.

## Integrity Rules

- Cite only sources actually returned by search this session; label anything else `[Background — not from search]` and exclude it from the findings count.
- Track queries sent / sources found / sources cited, and surface those counts in the closing audit block.
- Search sequentially, not all at once; on failure, retry once, then move on.
- After repeated failures across every source, stop, tell the user what's missing, and share whatever was collected — never deliver an empty report.

## Phase 0: Intake (2-4 questions, one at a time)

1. **Topic specificity** — state it in 1-2 sentences. "AI" or "tech" produces a vague survey; "self-hosted LLM deployment for small teams" produces a useful answer. Push back once on anything too broad ("What about AI — adoption, safety, capability, regulation, comparison? Pick an angle"); if still unnarrowed, proceed with an explicit "vague topic — survey level" caveat.
2. **Angle** — trend (what's accelerating/decelerating), sentiment (how people feel), problems (pain points/complaints), opportunities (gaps/unmet needs), or comparison (vs. alternatives). This determines which sources get weighted more heavily and how the synthesis is ranked. Default to trend if nothing else obviously fits.
3. **Time window** — 7 / 14 / 30 / 60 / 90 days (default 30). Shorter windows catch breaking conversation; longer ones catch sustained narrative shift.
4. **Platform scope** (asked only when the topic clearly makes some platforms low-value) — by default cover Reddit + Hacker News + open web. Reddit dominates sentiment, HN dominates technical critique, the open web dominates breadth — skip whatever doesn't fit the chosen angle.

Stop after question 4 and start research — never bundle questions, never reopen intake mid-run.

## Sources

**Reddit** — search both "top" and "new" sorts within the time window (top surfaces consensus, new catches breaking signal); for the top 3-5 posts by score, pull the top comments too, since discussion often carries more signal than the post itself.

**Hacker News** — search both stories and comments within the window. HN skews technical/builder — say so explicitly in the synthesis ("HN's voice is implementation-oriented; consumer sentiment will be under-represented here"). If a query returns nothing, broaden it by dropping uncommon terms before giving up.

**Open web** — three query angles: trusted-publisher coverage (major outlets, filtered to the time window), recent reviews, and honest-opinion sources (searches for "problems," "complaints," "worth it"). Fetch and read the actual pages rather than citing off search snippets alone — every claim needs to trace to a fetched source.

**X/Twitter** — optional and lowest priority, since its content overlaps heavily with Reddit/HN and mainly adds delta rather than primary signal. If there's no way to search it in this session, say so plainly in the report ("Skipped — no access to X/Twitter this session") rather than pretending to have data.

## Cross-Platform Synthesis

Once sources are gathered, look across all of them for:

1. **Consensus signals** — points where 3+ sources agree (highest confidence); cite the URLs backing each.
2. **Controversy signals** — points where sources disagree; note who says what.
3. **Pain points** — recurring complaints, especially from Reddit and the open web.
4. **Excitement signals** — recurring enthusiasm, especially from HN and X where available.
5. **Emerging trends** — mentions that show up only in the newest posts and are absent from older ones (compare "new" vs. "top" results).
6. **Gaps** — what's notably absent that you'd expect to find.

## Report Format

```
# [TOPIC] — Pulse (Last N Days)
Angle: [chosen angle]

## TL;DR
[2-3 sentences]

## Reddit
Top posts with score/comments/summary/link, then a narrative paragraph on what Reddit is saying.

## Hacker News
Notable stories with points/comments/summary/link, then a narrative paragraph (note the technical/builder bias).

## Web
Key sources with publication/takeaway/link, then a narrative paragraph.

## X/Twitter (if available)
Findings, or "Skipped — [reason]".

## Cross-Platform Patterns
Highest-confidence signals across sources.

## Key Takeaways
3-5 bullets.

## Content Angles (if applicable)
2-3 specific angles the data supports.

---
Audit: Queries sent: N (by source). Sources found: M. Sources cited: K.
```

## Why Sequential, Why Fetch-First

Firing every query for every source at once produces rate-limit failures and citation drift — running each source's searches in order, confirming each result before the next call, keeps the run reliable and makes it possible to say precisely how many queries were sent versus how many sources were actually usable. The fetch-before-cite rule for the open web exists for the same reason: a search-result snippet often mischaracterizes what a page actually says, so treat a snippet as a lead to open, not as a citable fact on its own.

## Error Handling

| Situation | Response |
|---|---|
| Topic too vague after one push-back | Proceed with an explicit "vague topic" caveat rather than refusing outright. |
| A source returns nothing | Broaden the query once; if still empty, note the gap rather than fabricating results. |
| A fetch times out | Use whatever loaded and mark that source "truncated." |
| X/Twitter unavailable | Include the section with "Skipped — [reason]"; never fake data. |
| Repeated failures across sources | Stop, return what was collected, note explicitly that the run stopped early — never deliver an empty report. |

## Anti-Patterns to Reject

- Starting any search before the topic is specific enough (question 1).
- Batching intake questions instead of asking one at a time.
- Citing background knowledge in the "cited" count.
- Fabricating sources to fill out a thin section.