# How the news-pulse Skill Performs Rapid News Attribution

> Discover how the news-pulse skill achieves rapid news attribution. Learn how four parallel scout agents analyze events, regulations, peers, and sentiment for quick insights.

- Repository: [Xbt Lin/ai-berkshire](https://github.com/xbtlin/ai-berkshire)
- Tags: how-to-guide
- Published: 2026-07-25

---

**The news-pulse skill achieves rapid news attribution by orchestrating four parallel scout agents that simultaneously scan company events, regulatory changes, industry peers, and market sentiment, synthesizing the findings into a prioritized attribution report within approximately ten minutes.**

The **news-pulse** skill is a core component of the `xbtlin/ai-berkshire` repository designed to explain stock price movements through evidence-based investigation. Unlike traditional sequential research workflows, this skill employs a multi-agent architecture defined in [[`skills/news-pulse.md`](https://github.com/xbtlin/ai-berkshire/blob/main/skills/news-pulse.md)](https://github.com/xbtlin/ai-berkshire/blob/main/skills/news-pulse.md) that executes four orthogonal investigations concurrently, delivering a concise attribution analysis significantly faster than manual methods.

## Parallel Agent Architecture

### Dynamic Team Creation

The workflow begins with a **TeamCreate** operation (lines 44-48) that establishes a dedicated team named `{ticker}-newspulse` to own the entire attribution process. This temporary team structure ensures isolated resource management and clean termination via **TeamDelete** (lines 22-23) after report generation. According to the source code, this pattern allows the skill to spin up specialized computational resources on demand without leaving persistent infrastructure.

### The Four Scout Agents

Once the team is established, the system launches four independent tasks using **TaskCreate** operations defined at lines 50-94. Each task runs in the background (`run_in_background: true`) and assigns a **general-purpose agent** to a specific domain:

- **Company-event scout**: Investigates corporate announcements, earnings reports, and management changes
- **Regulatory-watcher**: Monitors new regulations, fines, licensing changes, and government actions
- **Industry-peer tracker**: Analyzes competitor movements and sector-wide trends
- **Sentiment-tracker**: Assesses market narrative through social media and analyst opinions, optionally utilizing [[`tools/xueqiu_scraper.py`](https://github.com/xbtlin/ai-berkshire/blob/main/tools/xueqiu_scraper.py)](https://github.com/xbtlin/ai-berkshire/blob/main/tools/xueqiu_scraper.py) for Chinese "大 V" opinions

Agents communicate partial results back to the team lead through **TaskUpdate** and **SendMessage** operations (lines 44-46), enabling real-time aggregation without blocking the pipeline.

## Rapid Attribution Mechanisms

### Information-Availability Grading

The skill implements an **information-availability grading** system (lines 34-41) that assigns A, B, or C grades based on data richness for the target company. For **C-grade** companies with scarce data, the workflow switches to a "scan-only" mode at line 40, instantly concluding that no explanatory news is available rather than wasting resources on deep searches. This early-exit strategy ensures the system maintains its ~10-minute execution target even when facing data-poor scenarios.

### Dual-Source Verification and Parallel Fetching

Each agent executes a **dual-source verification** rule requiring at least two independent sources to confirm rumors before triggering resource-intensive operations. Only events surviving this filter initiate a **WebFetch** of the original announcement, keeping I/O to a minimum. Agents prepend date windows and "latest" keywords to their **WebSearch** queries (line 30), biasing results toward fresh content and reducing noise from historical data.

### Structured Synthesis and Persistence

The team lead merges the four dimensional reports using logic defined at lines 55-89, ranking events by relevance without requiring additional parsing. Agents return compact tables (date, event, source, relevance) that feed directly into the final synthesis. The completed report is written to `reports/{company}/{company}-news-{YYYYMMDD}.md` (lines 17-22) before the temporary team is dismantled.

## Implementation Details and Code Examples

### CLI Invocation

The skill is invoked through a CLI dispatcher that parses the company name, time window, and price movement context:

```bash
/news-pulse 腾讯 14 天 "跌 12%/3 天"

```

This command triggers the workflow defined in [`skills/news-pulse.md`](https://github.com/xbtlin/ai-berkshire/blob/main/skills/news-pulse.md), with an alternative interface available at [[`codex-skills/news-pulse/SKILL.md`](https://github.com/xbtlin/ai-berkshire/blob/main/codex-skills/news-pulse/SKILL.md)](https://github.com/xbtlin/ai-berkshire/blob/main/codex-skills/news-pulse/SKILL.md).

### Task Creation Configuration

The core parallel execution is established through YAML configuration:

```yaml
- TaskCreate:
    team_name: "tencent-newspulse"
    subagent_type: "general-purpose"
    run_in_background: true
    name: "company-event-scout"
    subject: "侦察 腾讯 近 14 天的公司本体事件"
    description: |
      1. 官方公告 …
      8. 输出按时间倒序的时间线表格

```

### Agent Prompt Template

Each agent receives a domain-specific prompt that includes the time window, stock movement context, and verification rules:

```text
你是 腾讯 新闻脉搏团队中的"监管与政策"，
负责侦察 监管 与政策 维度的最近 14 天事件。

时间窗口：2026‑04‑01 ~ 2026‑04‑15
股价异动背景：跌 12%/3 天
信息可得性等级：A

请完成任务 #2：regulatory-watcher

具体侦察要求：
1. 行业监管：所在行业的新规、罚款、整改、牌照变化
…
- 使用 WebSearch 进行时效性查询
- 对关键事件使用 WebFetch 精读原始来源
- 传言至少要 2 个独立来源

```

### Final Attribution Output

The synthesis produces a prioritized attribution table:

```markdown
#### 一句话归因

> 腾讯股价大幅下跌，主要因监管部门对其游戏业务的反垄断处罚（高）+ 市场情绪负面加剧（中）。

#### 完整事件时间线

| 日期       | 维度   | 事件                              | 来源 | 异动归因权重 |
|------------|--------|-----------------------------------|------|--------------|
| 2026‑04‑14 | 公司   | 腾讯发布 Q1 财报，营收低于预期   | …    | ⚪ 低 |
| 2026‑04‑13 | 监管   | 国家市场监管局发出《反垄断行政处罚决定书》 | …    | 🔴 高 |
| 2026‑04‑12 | 情绪   | 大V 在雪球上发布“腾讯将被处罚”预测 | …    | 🟡 中 |

```

## Summary

- The **news-pulse** skill creates a temporary team of four parallel agents via `TeamCreate` to investigate stock price movements from multiple angles simultaneously.
- **Information-availability grading** (A/B/C) and early-exit logic for C-grade companies ensure rapid completion even when data is scarce.
- **Dual-source verification** and selective **WebFetch** minimize I/O overhead while maintaining accuracy standards.
- Final attribution reports are persisted to `reports/{company}/{company}-news-{YYYYMMDD}.md` before the team is cleaned up via `TeamDelete`.

## Frequently Asked Questions

### How does the news-pulse skill verify the accuracy of news sources?

The skill implements a **dual-source verification** requirement where agents must confirm rumors through at least two independent sources before initiating a **WebFetch** operation. This rule, embedded in the agent prompt template at lines 17-33 of [`skills/news-pulse.md`](https://github.com/xbtlin/ai-berkshire/blob/main/skills/news-pulse.md), ensures that only validated information proceeds to the attribution synthesis stage, preventing false positives from single-source speculation.

### What happens when the news-pulse skill cannot find relevant news for a company?

When the **information-availability grading** system assigns a C-grade to a company (indicating scarce data availability), the workflow switches to a "scan-only" mode at line 40 of [`skills/news-pulse.md`](https://github.com/xbtlin/ai-berkshire/blob/main/skills/news-pulse.md). In this mode, the agent instantly concludes that no explanatory news is available rather than executing deep searches, allowing the system to complete within the target timeframe while acknowledging data limitations.

### Where does the news-pulse skill store its final attribution reports?

Upon completion, the team lead writes the final markdown report to the path `reports/{company}/{company}-news-{YYYYMMDD}.md` as specified at lines 17-22 of the skill definition. This standardized naming convention ensures reports are organized by company and date for easy retrieval, after which **TeamDelete** (lines 22-23) removes the temporary team to free resources.

### Which agents are involved in the news-pulse skill's parallel workflow?

The workflow creates four **general-purpose agents** through `TaskCreate` operations (lines 50-94): the **company-event scout**, **regulatory-watcher**, **industry-peer tracker**, and **sentiment-tracker**. Each runs as an independent background task, with the sentiment agent optionally leveraging [`tools/xueqiu_scraper.py`](https://github.com/xbtlin/ai-berkshire/blob/main/tools/xueqiu_scraper.py) for Chinese social media analysis.