| name | news-intel |
| description | Gather and rank symbol-relevant news intelligence from manual headlines or a local headline file. Use when user wants an explicit evidence-ingestion workflow for catalysts, risks, narrative tags, and freshness-aware headline triage. |
| argument-hint | ["symbol"] |
| allowed-tools | Bash(python3 *), Read, Glob, Grep |
News Intel
Gather ranked news intelligence for: $ARGUMENTS
Overview
- Implementation status: code-backed
- Local entry script:
<bundle-root>/news-intel/run.py
- Primary purpose: turn provided headline evidence into a ranked catalyst, risk, and narrative block for downstream research use
- Research layer: event-driven context (Stage 2: Data Collection & Quality Assurance, Stage 4: Feature Engineering & Signal Construction - Alternative data subset)
- Workflow stages: stage 2
Data Collection & Quality Assurance and stage 4 Feature Engineering & Signal Construction for event-oriented context
- Local executor guarantee: parse manual or file-backed headlines, rank them, tag catalysts and risks, and emit a compact report plus query-plan scaffolding
Use When
- The user wants symbol-relevant catalyst and risk context.
- The caller already has headlines and wants structured ranking rather than raw browsing.
- The user wants a compact narrative block before
analysis, stock-data, or strategy-chat.
- The user wants to understand how news events may impact technical or fundamental analysis.
- The user wants to identify potential catalysts for thesis formation or invalidation triggers.
Do Not Use When
- The user expects this skill to crawl the web on its own. The current local executor ranks provided headlines rather than performing broad autonomous search.
- The user wants a full one-symbol report. Use
analysis or market-brief.
- The user wants historical artifact review rather than fresh headline ingestion. Use
reports or analysis-history.
- The user wants real-time news monitoring or alerting. This skill processes provided headlines, not live feeds.
Inputs
- Normal case: one stock symbol.
- Optional
--headline "title | summary" repeated multiple times.
- Optional
--headline-file PATH for local text, JSON, or JSONL headline input.
- Optional
--company-name, --alias, --max-items.
- If
symbol is omitted, the skill may reuse last_symbol from the same execution context.
- Evidence contract:
- headline items are caller-supplied or file-supplied evidence
- source quality, timeliness, and completeness are not guaranteed by the local executor
- if no headlines are provided, the skill returns a minimal query-plan-oriented scaffold rather than pretending current news coverage exists
Execution
Step 1: Define news intelligence requirements
Before processing headlines, clarify the intelligence needs:
News intelligence objectives:
Event categories to identify:
- Earnings and financial: Results, guidance, analyst ratings, forecasts
- Corporate actions: M&A, restructuring, capital raising, dividends, buybacks
- Regulatory and policy: Approvals, investigations, policy changes, compliance
- Product and business: New products, contracts, partnerships, market expansion
- Management and governance: Leadership changes, strategy shifts, governance issues
- Market and sector: Industry trends, competitive dynamics, macro factors
- Technical and trading: Price moves, volume spikes, institutional activity
A-share specific events:
- Regulatory: CSRC approvals, delisting risk, ST/*ST designation changes
- Policy: Industry policy changes, subsidy programs, environmental regulations
- State ownership: SOE reforms, state asset transfers, mixed ownership reforms
- Capital markets: IPO/refinancing approvals, index inclusion/exclusion, Stock Connect changes
- Corporate governance: Shareholder disputes, related-party transactions, pledge releases
Step 2: Collect and validate headline evidence
Process provided headlines with quality assessment:
Headline collection:
- Accept manual headlines via
--headline flag
- Accept headline file via
--headline-file flag
- Parse headline format: "title | summary" or structured JSON/JSONL
- Extract metadata: timestamp, source, relevance score (if provided)
Headline validation:
Evidence quality assessment:
- Freshness: How old is the headline?
- Fresh: < 1 day old
- Recent: 1-7 days old
- Stale: > 7 days old
- Undated: Timestamp missing (treat as low quality)
- Source credibility:
- Official: Company announcements, exchange filings (highest credibility)
- Established media: Major financial news outlets (high credibility)
- Social media: Weibo, forums, blogs (medium credibility)
- User-supplied: Unknown provenance (low credibility)
- Completeness:
- Complete: Title, summary, timestamp, source all present
- Partial: Missing some metadata
- Minimal: Only title present
- Duplication:
- Unique: First report of this event
- Duplicate: Same event as another headline
- Update: New information on previously reported event
Step 3: Rank and categorize headlines
Apply systematic ranking and categorization:
Relevance scoring (0-100):
- Direct mention (80-100): Company name or ticker in title
- Indirect mention (50-79): Company in summary, or sector/industry mention
- Tangential (20-49): Related company, supply chain, or macro factor
- Irrelevant (0-19): No clear connection to symbol
Materiality scoring (0-100):
- High materiality (80-100): Likely to significantly impact price
- Earnings surprise, major M&A, regulatory approval/rejection
- Management change, fraud allegation, delisting risk
- Medium materiality (50-79): May impact price moderately
- Analyst rating change, contract win/loss, product launch
- Industry policy change, competitor news
- Low materiality (20-49): Minor impact expected
- Routine announcements, minor partnerships, general commentary
- Negligible (0-19): Unlikely to impact price
- Historical information, general industry trends
Sentiment classification:
- Bullish: Positive catalyst, likely to drive price up
- Bearish: Negative risk, likely to drive price down
- Neutral: Informational, no clear directional impact
- Mixed: Contains both positive and negative elements
Event timing classification:
- Immediate: Event just occurred, market may not have fully reacted
- Recent: Event occurred 1-7 days ago, likely already priced in
- Historical: Event occurred > 7 days ago, definitely priced in
- Forward-looking: Event expected in future (guidance, scheduled events)
Confidence scoring (0-100):
- High confidence (80-100): Official source, complete information, clear impact
- Medium confidence (50-79): Established media, mostly complete, probable impact
- Low confidence (20-49): Unverified source, incomplete information, unclear impact
- Speculative (0-19): Rumor, no source, highly uncertain
Step 4: Extract catalysts and risks
Identify specific catalysts and risks from ranked headlines:
Catalyst extraction:
For each bullish headline with materiality > 50:
- Catalyst type: Earnings beat, M&A, approval, contract, etc.
- Catalyst description: Brief summary of the positive event
- Expected impact: How this could drive price up
- Timing: When impact may materialize
- Confidence: How certain is this catalyst
- Evidence: Which headline(s) support this catalyst
Risk extraction:
For each bearish headline with materiality > 50:
- Risk type: Earnings miss, investigation, competition, regulation, etc.
- Risk description: Brief summary of the negative event
- Expected impact: How this could drive price down
- Timing: When impact may materialize
- Confidence: How certain is this risk
- Evidence: Which headline(s) support this risk
Narrative theme extraction:
Identify recurring themes across headlines:
- Growth narrative: Expansion, market share gains, new products
- Quality narrative: Margin improvement, efficiency gains, competitive moats
- Value narrative: Undervaluation, asset sales, shareholder returns
- Risk narrative: Regulatory pressure, competition, macro headwinds
- Sentiment narrative: Market enthusiasm, skepticism, controversy
Step 5: Assess evidence quality and gaps
Evaluate the overall quality of news intelligence:
Evidence quality summary:
- Total headlines processed
- Freshness distribution (fresh/recent/stale/undated)
- Source distribution (official/media/social/user-supplied)
- Completeness distribution (complete/partial/minimal)
- Duplication rate (% of headlines that are duplicates)
Evidence gaps:
- Missing event types: Which important event categories have no coverage?
- One-sided coverage: Are headlines predominantly bullish or bearish?
- Stale information: Is most information > 7 days old?
- Low credibility: Are most headlines from unverified sources?
- Sparse coverage: Are there too few headlines to draw conclusions?
Quality flags:
Step 6: Synthesize news intelligence report
Organize findings into structured report:
Part 1: Intelligence Summary
- Symbol and company name
- Total headlines processed
- Date range of headlines
- Overall sentiment (bullish/bearish/neutral/mixed)
- Evidence quality score (0-100)
Part 2: Key Catalysts
For each catalyst (ranked by materiality × confidence):
- Catalyst type and description
- Expected impact and timing
- Confidence level
- Supporting evidence (headline titles)
Part 3: Key Risks
For each risk (ranked by materiality × confidence):
- Risk type and description
- Expected impact and timing
- Confidence level
- Supporting evidence (headline titles)
Part 4: Narrative Themes
- Dominant narratives across headlines
- Sentiment trends (improving/deteriorating/stable)
- Market positioning (how company is perceived)
- Sector context (industry trends affecting company)
Part 5: Ranked Headlines
Top 10-20 headlines by relevance × materiality:
- Title and summary
- Timestamp and source
- Relevance, materiality, sentiment scores
- Confidence level
Part 6: Evidence Quality Assessment
- Freshness summary
- Source credibility summary
- Completeness summary
- Duplication summary
- Quality flags (if any)
Part 7: Evidence Gaps and Limitations
- Missing event types
- One-sided coverage concerns
- Stale information concerns
- Low credibility concerns
- Sparse coverage concerns
- Recommended additional research
Part 8: Query Plan (Optional)
If evidence is insufficient, suggest queries for additional research:
- Specific events to investigate
- Sources to check (exchange filings, company website, etc.)
- Time periods to focus on
- Related companies or sectors to monitor
Step 7: Run the local executor
python3 <bundle-root>/news-intel/run.py <symbol> [--headline "标题 | 摘要"] [--headline-file PATH]
Step 8: Deliver bounded intelligence block
When presenting results, maintain strict boundaries:
Evidence provenance:
- State that headlines are user-supplied or file-supplied evidence
- Identify source for each major catalyst or risk claim
- Distinguish official sources from media reports from speculation
Ranking transparency:
- Explain that scores are heuristic rankings, not truth scores
- State that relevance/materiality/confidence are model outputs, not guarantees
- Acknowledge that ranking cannot assess information not provided
Freshness and duplication:
- Highlight when most relevant items are stale (> 7 days old)
- Flag when multiple headlines report the same event
- Note when evidence is sparse or one-sided
Inference boundaries:
- Separate directly supplied evidence from derived insights
- Label narrative themes as interpretive synthesis, not facts
- Qualify catalyst/risk claims based on evidence quality
Integration with analysis:
- Catalysts can support thesis formation in
analysis skill
- Risks can inform disconfirming evidence in
analysis skill
- Narrative themes can inform variant view in
analysis skill
- Evidence gaps should narrow claim strength in downstream memos
Output Contract
- Minimum local executor output: human-readable text beginning with
# <symbol> News Intel.
- Possible sections:
Catalysts, Risks, Narratives, Ranked Items, and Query Plan.
- Side effects: updates session memory for the current execution context.
- Caller-facing delivery standard:
- Eight-part structure: Intelligence summary, key catalysts, key risks, narrative themes, ranked headlines, evidence quality, evidence gaps, query plan (if needed)
- Evidence provenance: Label headline basis as supplied evidence, not independently verified truth
- Ranking transparency: Treat confidence and relevance scores as heuristic ranking outputs, not completeness or truth scores
- Freshness disclosure: Surface freshness and duplication profile when it materially affects conclusions
- Quality assessment: Provide evidence quality score and flag quality concerns
- Inference boundaries: Separate supplied evidence, ranking output, and narrative interpretation
- Narrow claims when evidence is thin: If evidence is sparse, stale, or one-sided, qualify conclusions accordingly
- Integration guidance: Explain how catalysts/risks feed into thesis formation, disconfirming evidence, and invalidation triggers
- No fabricated certainty: Incomplete evidence should degrade conclusion strength, not be hidden
- Non-guarantees:
- No web crawling guarantee (processes provided headlines only)
- No durable report artifact unless another skill persists the result
- No real-time monitoring or alerting capability
Failure Handling
- Parse and argument errors: non-zero exit with a readable
命令错误 message.
- Missing or unreadable headline files: readable
执行失败: text.
- Missing headlines: returns a reduced report with query plan suggestions.
- If evidence is incomplete, that should degrade the conclusion rather than cause fabricated certainty.
- Malformed headline format: skip malformed entries and report count of skipped items.
- Duplicate detection failures: proceed with ranking but flag potential duplication issues.
Key Rules
- Distinguish provided evidence from inferred narrative.
- Treat ranking as deterministic scoring over supplied items, not as a guarantee of external completeness.
- Prefer concise, source-labeled headline payloads over noisy dumps.
- When this skill feeds
analysis, preserve provenance for each major catalyst or risk claim.
- Evidence quality assessment is mandatory. Always evaluate freshness, source credibility, completeness, duplication.
- Quality flags must be surfaced. If evidence is insufficient, stale, low-credibility, one-sided, or duplicate-heavy, state it explicitly.
- Confidence scores must reflect evidence quality. Low-quality evidence = low confidence, regardless of headline content.
- Narrative themes are interpretive synthesis. Label them as such, not as facts.
- Catalyst and risk claims must be qualified. State expected impact, timing, and confidence level.
- Evidence gaps must be acknowledged. Missing event types, one-sided coverage, sparse information.
- Integration with downstream skills must be explicit. How do catalysts/risks inform thesis, disconfirming evidence, invalidation?
- No web crawling or autonomous search. This skill processes provided headlines only.
Composition
- Often feeds
analysis, stock-data, or strategy-chat.
- Shares research helpers with the broader packet-building logic, but it is not a substitute for full research verification.
- Catalysts feed into thesis formation and key drivers in
analysis skill.
- Risks feed into disconfirming evidence and invalidation conditions in
analysis skill.
- Narrative themes feed into variant view and market positioning in
analysis skill.
- Evidence gaps inform evidence sufficiency assessment in
analysis skill.
- Can be combined with
fundamental-context for comprehensive non-price context.
- Should be rerun when new material events occur to update catalyst/risk assessment.
