deep-reasoning-agent
4
总安装量
4
周安装量
#51427
全站排名
安装命令
npx skills add https://github.com/psh355q-ui/szdi57465yt --skill deep-reasoning-agent
Agent 安装分布
claude-code
3
windsurf
2
trae
2
opencode
2
codex
2
antigravity
2
Skill 文档
Deep Reasoning Agent – 3ë¨ê³ ì¬ì¸µ ë¶ì
Role
/deep-reasoning íì´ì§ìì ë´ì¤ 기ì¬ë¥¼ **3ë¨ê³ Chain of Thought (CoT)**ë¡ ì¬ì¸µ ë¶ìí©ëë¤. ìëë³´ë¤ ê¹ì´ë¥¼ ì°ì ìí©ëë¤.
Core Capabilities
1. Three-Stage Chain of Thought
Stage 1: Direct Impact (ì§ì ìí¥)
Goal: ë´ì¤ê° í´ë¹ 기ì
ì 미ì¹ë ì¦ê°ì ì´ê³ ì§ì ì ì¸ ìí¥ ë¶ì
Questions:
- ì´ ë´ì¤ë 무ìì ë§íëê°?
- íì¬ì ì´ë¤ ë¶ë¶ì ìí¥ì 주ëê°?
- ì¬ë¬´ì ìí¥ì ì¼ë§ë ëëê°?
- ìê° íë ìì? (ì¦ì vs ì¥ê¸°)
Output:
- ì§ì ìí¥ ìì½
- Impact Score (0-1)
- Timeframe (immediate, short-term, long-term)
Example:
News: "FDA approves XYZ cancer drug"
Stage 1 Analysis:
- ì§ì ìí¥: ì ì½ í매 ì¹ì¸ â ë§¤ì¶ ì¦ê°
- ìì 매ì¶: ì°ê° $5B (analyst estimates)
- ìí¥ í¬ê¸°: VERY_HIGH (0.9)
- Timeframe: Short-term (6-12 months to ramp up)
Stage 2: Secondary Effects (2ì°¨ íê¸ í¨ê³¼)
Goal: ê³µê¸ë§, ê²½ìì¬, ê´ë ¨ ì°ì
ì 미ì¹ë ê°ì ìí¥ ë¶ì
Questions:
- ê²½ìì¬ë ì´ë¤ ìí¥ì ë°ëê°?
- ê³µê¸ë§(upstream/downstream)ì?
- ê·ì íê²½ ë³í?
- ìì¥ ì ì ì¨ ë³í?
Output:
- ìí¥ ë°ë í°ì»¤ 리ì¤í¸
- ê° í°ì»¤ë³ ìí¥ ë°©í¥ (positive/negative)
- ì°ì
ì ì²´ ìí¥
Example:
News: "Tesla announces 20% price cut"
Stage 2 Analysis:
- ê²½ìì¬ ìí¥:
* GM, F: NEGATIVE (ê°ê²© ê²½ì ìë°)
* RIVN, LCID: VERY_NEGATIVE (ìê·ëª¨ ì
ì²´, ê°ê²© ëì ì´ë ¤ì)
- ê³µê¸ë§:
* Battery suppliers (PANW, LGìëì§ì루ì
): NEGATIVE (주문ë ê°ì ì°ë ¤)
* Charging network (CHPT): NEUTRAL (볼륨 ì¦ê° ê°ë¥)
- ì°ì
ìí¥: ì 기차 ê°ê²© íë½ ìë° â ë³´ê¸ ê°ìí
Stage 3: Final Conclusion (ìµì¢ ê²°ë¡ )
Goal: Stage 1 + Stage 2 ì¢
í©íì¬ í¬ì ê²°ì ë° ì ëµ ì립
Questions:
- ì¢
í© íë¨: BUY/SELL/HOLD?
- ìê°ëë³ ì ëµ?
- 주ì 리ì¤í¬ë?
- ëì ìë리ì¤ë?
Output:
- Action (BUY/SELL/HOLD)
- Confidence (0-1)
- Short-term vs Long-term ì ëµ
- Risk Factors
- Alternative Scenarios
Example:
Conclusion:
- Action: BUY (the drug company)
- Confidence: 0.85
- Short-term (1-3 months): STRONG BUY (FDA ì¹ì¸ 모ë©í
)
- Long-term (6-12 months): BUY (ë§¤ì¶ ë³¸ê²©í)
- Risks:
* ë³´í coverage ë¶íì¤ì±
* ê²½ì ì½ë¬¼ ê°ë° ê°ë¥ì±
- Alternative Scenario:
* IF insurance rejects coverage â ì£¼ê° -15%
* IF competitor announces similar drug â ì£¼ê° -10%
2. Related Tickers Analysis
def find_related_tickers(news_article: NewsArticle) -> List[Dict]:
"""Find all tickers affected by the news"""
related = []
# Primary ticker (mentioned in news)
primary = news_article.ticker
# Competitors (same sector)
competitors = get_competitors(primary)
# Supply chain
suppliers = get_suppliers(primary)
customers = get_customers(primary)
# Industry ETFs
etfs = get_related_etfs(primary)
return {
"primary": primary,
"competitors": competitors,
"suppliers": suppliers,
"customers": customers,
"etfs": etfs
}
3. Impact Quantification
def quantify_impact(
news_type: str,
magnitude: str,
company_size: str
) -> Dict:
"""Estimate price impact"""
# Base impact by news type
BASE_IMPACT = {
"fda_approval": 0.15, # +15% average
"earnings_beat": 0.05, # +5%
"merger": 0.20, # +20%
"lawsuit": -0.10, # -10%
"ceo_departure": -0.08 # -8%
}
# Magnitude multiplier
MAGNITUDE = {
"small": 0.5,
"medium": 1.0,
"large": 1.5
}
# Company size adjustment
SIZE_ADJ = {
"large_cap": 0.7, # Less volatile
"mid_cap": 1.0,
"small_cap": 1.3 # More volatile
}
base = BASE_IMPACT.get(news_type, 0.05)
mag = MAGNITUDE.get(magnitude, 1.0)
size = SIZE_ADJ.get(company_size, 1.0)
estimated_impact = base * mag * size
return {
"estimated_price_change_pct": estimated_impact,
"confidence": 0.6, # Historical accuracy
"timeframe": "1-3 months"
}
Decision Framework
Step 1: Receive News Article
- news_id: 123
- ticker: MRNA
- headline: "FDA Approves Cancer Vaccine"
- content: [full article]
Step 2: Stage 1 Analysis (Direct Impact)
prompt_stage1 = f"""
Analyze the DIRECT impact of this news on {ticker}:
News: {headline}
{content}
Answer:
1. What happened?
2. How does it affect the company's revenue?
3. What is the financial impact?
4. When will impact be felt?
"""
stage1_result = call_ai(prompt_stage1)
Step 3: Stage 2 Analysis (Secondary Effects)
# Find related tickers
related = find_related_tickers(ticker)
prompt_stage2 = f"""
Stage 1 conclusion: {stage1_result}
Now analyze SECONDARY effects:
Competitors: {related['competitors']}
Suppliers: {related['suppliers']}
Answer:
1. How do competitors react?
2. Supply chain impact?
3. Industry-wide changes?
"""
stage2_result = call_ai(prompt_stage2)
Step 4: Stage 3 Conclusion
prompt_stage3 = f"""
Stage 1: {stage1_result}
Stage 2: {stage2_result}
Provide FINAL trading decision:
1. BUY/SELL/HOLD?
2. Short-term vs Long-term strategy?
3. Key risks?
4. Alternative scenarios?
"""
stage3_result = call_ai(prompt_stage3)
Step 5: Generate Trading Signal
IF stage3_result.action == "BUY":
create_trading_signal(
ticker=ticker,
action="BUY",
source="deep_reasoning",
confidence=stage3_result.confidence,
metadata={
"news_id": news_id,
"stage1": stage1_result,
"stage2": stage2_result,
"stage3": stage3_result
}
)
Output Format
{
"news_id": 123,
"ticker": "MRNA",
"headline": "FDA Approves Moderna Cancer Vaccine",
"analysis_timestamp": "2025-12-21T13:00:00Z",
"analysis_duration_sec": 28,
"stage1_direct_impact": {
"summary": "FDA ì¹ì¸ì¼ë¡ Modernaì ì ë°±ì ì´ ìì¥ ì§ì
. ì°ê° ë§¤ì¶ $5B ì¶ì (ë¶ìê° ì»¨ì¼ìì¤). íì¬ ì´ ë§¤ì¶ì ~40% ì¦ê° ìì.",
"impact_score": 0.9,
"impact_level": "VERY_HIGH",
"timeframe": "short_term",
"financial_estimates": {
"annual_revenue_potential": 5000000000,
"margin_estimate": 0.65,
"market_exclusivity_years": 7
},
"reasoning": "ì ì½ ì¹ì¸ì ì¦ê°ì ì¸ ë§¤ì¶ ê¸°í ì°½ì¶. Modernaë mRNA íë«í¼ì ì
ì¦ë 리ëë¡ ë¹ ë¥¸ ìì©í ê°ë¥."
},
"stage2_secondary_effects": {
"summary": "ê²½ì ì ì½ì¬(PFE, MRCK)ë ì ë°±ì ê²½ì ì¬í. mRNA ê³µê¸ë§(LNP suppliers) ìí. í¬ì¤ì¼ì´ ì¹í° ì ì²´ ê¸ì ì .",
"affected_tickers": [
{
"ticker": "PFE",
"relationship": "competitor",
"impact": "NEGATIVE",
"impact_score": -0.3,
"reasoning": "ìì¥ ì ì ì¨ ìí, ê²½ì ì¬í"
},
{
"ticker": "MRCK",
"relationship": "competitor",
"impact": "NEGATIVE",
"impact_score": -0.2,
"reasoning": "ì ì¹ë£ ìì¥ ê²½ì ì¦ê°"
},
{
"ticker": "NVAX",
"relationship": "competitor",
"impact": "NEUTRAL",
"impact_score": 0.1,
"reasoning": "ë¤ë¥¸ ì§ë³ í¬ì»¤ì¤, ì§ì ê²½ì ì ì"
},
{
"ticker": "XLV",
"relationship": "sector_etf",
"impact": "POSITIVE",
"impact_score": 0.2,
"reasoning": "í¬ì¤ì¼ì´ íì ê¸ì ì ì í¸"
}
],
"industry_impact": "mRNA 기ì ì
ì§ ê°í, ì ì¹ë£ í¨ë¬ë¤ì ì í 기ëê°",
"supply_chain_effects": "LNP(Lipid Nanoparticle) ìì ì¦ê°, CDMO ìí"
},
"stage3_conclusion": {
"action": "BUY",
"confidence": 0.85,
"reasoning": "Stage 1 ë§¤ì° ê¸ì ì ì§ì ìí¥ + Stage 2 ê²½ìì¬ ì½ì¸ë MRNAì ê²½ì ì°ì ê°í. ë¨ê¸° 모ë©í
+ ì¥ê¸° íëë©í¸ 모ë ìí¸.",
"time_horizon_strategy": {
"short_term_1_3_months": {
"action": "STRONG_BUY",
"confidence": 0.90,
"rationale": "FDA ì¹ì¸ 모ë©í
, 미ëì´ ì£¼ëª©, ê¸°ê´ ë§¤ì ìì",
"target_price": 185.00,
"expected_return": 0.18
},
"medium_term_3_6_months": {
"action": "BUY",
"confidence": 0.80,
"rationale": "ìì©í ì§í, ì´ê¸° ë§¤ì¶ ë°ì´í° ê³µê°",
"target_price": 200.00,
"expected_return": 0.28
},
"long_term_6_12_months": {
"action": "HOLD_OR_BUY",
"confidence": 0.70,
"rationale": "ë§¤ì¶ ë³¸ê²©í, íì§ë§ ê²½ì ì½ë¬¼ ì¶í ê°ë¥ì±",
"target_price": 210.00,
"expected_return": 0.34
}
},
"risk_factors": [
{
"risk": "ë³´í coverage ë¶íì¤ì±",
"probability": 0.30,
"impact": "HIGH",
"mitigation": "FDA ì¹ì¸ í ë³´íì¬ íì 주ì"
},
{
"risk": "ê²½ì ì½ë¬¼ íì´íë¼ì¸",
"probability": 0.40,
"impact": "MEDIUM",
"mitigation": "PFE, MRCK ìì ë°ì´í° 모ëí°ë§"
},
{
"risk": "ë¶ìì© ë³´ê³ ",
"probability": 0.15,
"impact": "VERY_HIGH",
"mitigation": "ì´ê¸° phase 4 ë°ì´í° 주ì"
}
],
"alternative_scenarios": [
{
"scenario": "ë³´í coverage ê±°ë¶",
"probability": 0.20,
"price_impact": -0.15,
"action_change": "HOLD â SELL"
},
{
"scenario": "ê²½ìì¬ ì ì¬ ì½ë¬¼ ì¹ì¸ (6ê°ì ë´)",
"probability": 0.25,
"price_impact": -0.10,
"action_change": "BUY â HOLD"
},
{
"scenario": "ì´ê¸° ë§¤ì¶ ê¸°ëì¹ ì´ê³¼",
"probability": 0.35,
"price_impact": +0.20,
"action_change": "BUY â STRONG_BUY"
}
],
"key_catalysts": [
"ë³´í coverage ë°í (positive)",
"ìì ì¶ê° ë°ì´í° (efficacy íì¸)",
"êµì ì¹ì¸ (EU, Japan)"
]
},
"trading_signal_generated": true,
"signal_id": "SIG-20251221-045"
}
Examples
Example 1: FDA Approval (ì ìì)
Example 2: Negative News (Lawsuit)
News: "Tesla faces $10B lawsuit over Autopilot defects"
Stage 1:
- ì§ì ìí¥: ë²ì ë¹ì© + ë¸ëë ì´ë¯¸ì§ í격
- ì¬ë¬´ ìí¥: ìµì
$10B (unlikely), íì¤ì $1-2B settlement
- Impact Score: 0.6 (MEDIUM-HIGH)
Stage 2:
- ê²½ìì¬: GM, F â POSITIVE (Tesla ì½ì ë¶ê°)
- ê·ì : ìì¨ì£¼í ê·ì ê°í ê°ë¥ì± â ì ì²´ ì¹í° NEGATIVE
- ê³µê¸ë§: Neutral
Stage 3:
- Action: SELL (short-term), HOLD (long-term)
- Confidence: 0.70
- Short-term: ë¶ì ì sentiment ì£¼ê° ìë° ìì
- Long-term: Tesla ë¸ëë íìë¡ íë³µ ê°ë¥
Guidelines
Do’s â
- ê¹ì´ ì°ì : ìëë³´ë¤ ì íì±ê³¼ ê¹ì´
- 3ë¨ê³ ì격 ì¤ì: ê° Stage ëª íí 구ë¶
- Related Tickers í¬í¨: 2ì°¨ ìí¥ ë¶ì íì
- ìëë¦¬ì¤ ë¶ì: Alternative scenarios ì ì
Don’ts â
- ë¨ê³ ê±´ëë°ê¸° ê¸ì§
- íë©´ì ë¶ì ê¸ì§ (Quick Analyzerì ì°¨ë³í)
- Related tickers ëë½ ê¸ì§
- Risk factors ìëµ ê¸ì§
Integration
API Endpoint
@router.post("/api/deep-reasoning/analyze")
async def deep_reasoning_analysis(news_id: int, db: Session):
"""Deep 3-stage analysis for a news article"""
# Get news article
news = db.query(NewsArticle).filter_by(id=news_id).first()
if not news:
raise HTTPException(404, "News not found")
# Run Deep Reasoning Agent
agent = DeepReasoningAgent()
result = await agent.execute({
'news_id': news_id,
'ticker': news.ticker,
'headline': news.headline,
'content': news.content
})
# Generate trading signal
if result['stage3_conclusion']['action'] in ['BUY', 'SELL']:
create_trading_signal(
ticker=news.ticker,
action=result['stage3_conclusion']['action'],
confidence=result['stage3_conclusion']['confidence'],
source='deep_reasoning',
reasoning=result['stage3_conclusion']['reasoning'],
metadata=result
)
return result
Performance Metrics
- Analysis Time: íê· 20-30ì´ (ê¹ì´ ì°ì )
- Accuracy: 목í > 75% (Quick Analyzer 60% ëë¹ ëì)
- Related Tickers Recall: > 90% (주ì ìí¥ í°ì»¤ í¬ì°©)
- User Satisfaction: > 4.5/5 (ê¹ì´ê°)
Comparison
| Agent | Speed | Depth | Accuracy | Use Case |
|---|---|---|---|---|
| Quick Analyzer | 5ì´ | â | 60% | ë¹ ë¥¸ íì¸ |
| Deep Reasoning | 30ì´ | âââ | 75% | ì¤ìí ê²°ì |
| War Room | 15ì´ | ââ | 65% | í©ì ê¸°ë° |
Version History
- v1.0 (2025-12-21): Initial release with 3-stage Chain of Thought methodology