Using the news-pulse skill for stock price movement attribution in AI Berkshire

The news-pulse skill delivers a 10-15 minute rapid attribution analysis by orchestrating four parallel scout agents to investigate what caused a stock's price swing.

AI Berkshire is a layered framework that transforms single LLM prompts into comprehensive research workflows. The news-pulse skill provides automated stock price movement attribution when equities experience significant volatility. According to the source code in skills/news-pulse.md, this skill follows a strict three-tier architecture designed to maximize information throughput while maintaining analytical rigor.

Architecture of the news-pulse skill

The skill implements a deterministic multi-agent pipeline through three distinct layers.

Skill layer

The skill layer is defined declaratively in skills/news-pulse.md. This Markdown file specifies the command interface, required arguments (company name, time window, price swing percentage), and the complete execution flow. The file outlines how parameters are validated and an information-availability rating (A/B/C) is computed before execution begins.

Agent layer

The agent layer spawns a team-lead agent and four parallel "scout" agents: company-event, regulatory-watcher, industry-peer, and sentiment-tracker. Each agent operates independently with custom prompt templates defined in the "第四步：创建 4 个侦察任务" section of skills/news-pulse.md. This parallelization achieves four-times the search volume of single-agent workflows.

Tool layer

The tool layer provides agents with WebSearch, WebFetch, TaskCreate, TaskUpdate, SendMessage, and TeamCreate capabilities. Numeric data validation is enforced through tools/financial_rigor.py, which guarantees cross-validation and precise decimal calculations for any financial figures gathered during research.

Step-by-step workflow execution

The news-pulse skill executes a six-phase deterministic pipeline:

1. Parameter validation and rating calculation. Inputs including company name, time window, price swing magnitude, and focus area are validated. An information-availability rating (A/B/C) is computed to determine research depth.

2. Team initialization. The system creates a dedicated team using TeamCreate with the naming convention {company}-newspulse.

3. Parallel scout deployment. Four TaskCreate calls launch the scouting agents simultaneously, each receiving specialized prompt templates from the "每个 Agent 的 prompt 模板" block in skills/news-pulse.md.

4. Independent source verification. Agents execute in parallel, using WebFetch to retrieve primary sources and performing independent verification. Each outputs a compact core-findings list plus a full timeline table.

5. Synthesis and attribution. The team-lead agent aggregates findings into a one-sentence attribution, an attribution table quantifying contributor weights, a nature-of-movement judgement (value event / sentiment / mixed / unknown), and an action-recommendation checklist.

6. Report generation and cleanup. The final markdown report is written to reports/{company}/{company}-news-{YYYYMMDD}.md, and the temporary team is removed via TeamDelete.

For C-level (scarce-information) stocks, the skill automatically enters "扫盲模式" (basic education mode) and explicitly emits "真因不明" (true cause unknown) when data is insufficient.

How to invoke the news-pulse skill

Users can trigger the skill through three interfaces.

Claude Code slash command

The simplest invocation uses the native slash command:

/news-pulse 腾讯

This generates a markdown report at reports/腾讯/腾讯-news-20260702.md with the current date.

Codex prompt interface

For Codex users, the wrapper file codex-prompts/news-pulse.md forwards requests identically:

Use news-pulse to analyse 拼多多 跌12% 一周内

Programmatic JSON payload

For backend integration, the system accepts structured payloads:

{
  "skill": "news-pulse",
  "arguments": {
    "company": "阿里巴巴",
    "window_days": 14,
    "price_move": "涨 8% 5 天"
  }
}

The backend converts this into the standard TeamCreate, TaskCreate, and TaskUpdate sequences defined in the skill file.

Understanding the output format

The generated markdown report follows a standardized structure designed for rapid decision-making. The output includes a one-sentence summary, quantitative attribution table, confidence assessments, and actionable recommendations.

A typical attribution table appears as follows:

#### 一句话归因

这次 8% 涨幅主要由行业 Beta + 新品发布驱动，价值事件贡献约 20%，情绪贡献约 80%。

#### 异动归因表

| 候选解释       | 估算贡献 | 置信度 |
|----------------|----------|--------|
| 新品发布（双11预热） | +5%      | 高 |
| 行业 Beta (电商整体复苏) | +2%      | 中 |
| 市场情绪（大V看好） | +1%      | 高 |
| 基本面未出现负面 | 0%      | 极高（排除） |

Summary

• The news-pulse skill provides automated stock price movement attribution through a deterministic six-step multi-agent workflow defined in skills/news-pulse.md.
• Four parallel scout agents (company-event, regulatory-watcher, industry-peer, sentiment-tracker) execute simultaneously to maximize information gathering within a 10-15 minute window.
• The information-availability rating (A/B/C) automatically adjusts research depth, with C-level stocks entering a specialized "basic education mode" when data is scarce.
• Financial rigor is enforced through tools/financial_rigor.py, ensuring all numeric claims are cross-validated before inclusion in the final report.
• Output is generated as a structured markdown file in reports/{company}/{company}-news-{YYYYMMDD}.md, containing one-sentence attribution, contribution tables, and investment thesis recommendations.

Frequently Asked Questions

How does the news-pulse skill handle low-information stocks?

When the skill computes a C-level information-availability rating for a stock, it automatically switches to "扫盲模式" (basic education mode). In this mode, the scout agents focus on gathering foundational company data rather than specific event attribution. If insufficient data exists to determine causation, the system explicitly emits "真因不明" (true cause unknown) rather than generating speculative conclusions.

What is the difference between the skill layer and the agent layer in AI Berkshire?

The skill layer consists of declarative Markdown files like skills/news-pulse.md that define command syntax, argument schemas, and workflow orchestration. The agent layer represents the runtime instantiation of LLM agents (the team-lead and four scouts) that execute the actual research tasks. The skill layer is static configuration; the agent layer is dynamic execution.

Can I customize the scout agents or add additional data sources?

While the default configuration in skills/news-pulse.md specifies four fixed scout types with predefined prompt templates, the architecture supports modification at the skill definition level. Users can edit the Markdown skill file to adjust agent prompts or add additional TaskCreate calls for supplementary scouts. However, the current implementation treats the four-agent pattern as the optimized balance between coverage and latency for the 10-15 minute attribution window.

Where are the final reports stored and how long do they persist?

The system writes finalized attribution reports to reports/{company}/{company}-news-{YYYYMMDD}.md using the company name and execution date. The temporary agent team created via TeamCreate is automatically torn down via TeamDelete immediately after report generation, but the markdown files persist in the reports directory for historical reference and audit trails.

{
  "mcpServers": {
    "instagit": {
      "command": "npx",
      "args": ["-y", "instagit@latest"]
    }
  }
}