alpha-research

/alpha-research

Purpose

/alpha-research 是唯一主入口 superpower。

它负责组织流程、动态追问、做研究工具栈 preflight、记录决策、维护输出结构，并把研究推进到可审计的本地报告。

最终产物是研究报告。可以给出有逻辑、有数据、有引用的公司/候选标的研究推荐，但不能输出直接交易指令、目标价、仓位建议或收益承诺。

Bundled files

• Templates:
  • templates/run-log.md.
  • templates/source-gate.md.
  • templates/tool-status.md.
• Resources:
  • resources/research-tool-stack.md.
  • resources/source-profiles.md.
  • resources/source-recipes.md.
  • resources/tool-install-sources.md.
  • resources/data-providers/*.md.

Use resources/research-tool-stack.md as the authoritative source for setup modes, channel status, source tiers, market-specific sources, minimum acceptable evidence, hard source gate, and quality ceiling rules. Use resources/source-profiles.md to choose the required source mix for the research object. Use resources/source-recipes.md to run the acquisition sequence. Use resources/tool-install-sources.md as the canonical install-source catalog. Use resources/data-providers/*.md for source-specific playbooks.

Default output root

Use this runtime artifact structure:

alpha-research-output/{topic-slug}/
  run-log.md
  research-brief.md
  source-map.md
  chain-trace.md
  thesis-audit.md
  report-quality.md
  report-source.md  # optional Markdown intermediate
  report.html       # final report

Final report format is report.html. Markdown files are process artifacts or optional source drafts. The default report style is Dense Calm + ECharts: high information density, calm visual surface, compact tables, plain-language explanations, and ECharts SVG figures.

Interview mode

Start like grill-me:

• Ask one question at a time.
• Provide a recommended answer for each question.
• Ask dynamically based on the user's previous answer.
• Do not ask what can already be inferred from the conversation or existing artifacts.
• Record decisions in run-log.md.

Useful decision branches:

• Topic.
• Research object: trend, industry, company, material, technology, or narrative.
• Current belief.
• Biggest fear of being wrong.
• Market scope: US, HK, A-share, China macro, global, or custom.
• Research mode: normal or anti-hype.
• Output language.
• Final report format: HTML by default and required.
• Optional Markdown source artifact only if helpful.
• Existing materials to include.
• Time horizon: short-term event, 1–3 year trend, or 3–5 year trend.

Fixed defaults:

• Evidence standard: strict.
• Source collection: enabled by default.
• Candidate/company recommendations: allowed only when backed by source-gated evidence, explicit logic, risks, counterarguments, and failure criteria.
• Local artifacts: default output under alpha-research-output/{topic-slug}/.
• Reader quality: default to an intelligent non-specialist investor/research reader. Explain jargon, expand abbreviations, and avoid unexplained English labels.
• Final report quality: the reader should learn the business mechanism, the data model, and both positive/negative research views. Black-box jargon summaries are not acceptable.

Setup preflight

/alpha-research owns setup preflight. Users do not need a separate setup skill.

Research preflight

Before synthesis and final report generation, choose the relevant profile(s) from resources/source-profiles.md, follow the matching recipe(s) from resources/source-recipes.md, then check all 11 information channels from resources/research-tool-stack.md and identify the concrete tool/source path for each material channel:

1. Web search.
2. Web / PDF fetch.
3. Filing / disclosure.
4. Company IR.
5. Financial data.
6. Academic / technical.
7. Patent / standards.
8. News / event.
9. Social / narrative.
10. Browser automation.
11. Alternative data leads.

Apply the hard source gate and quality ceiling rules from resources/research-tool-stack.md.

Alpha Research cannot rely on web search alone when source-specific channels are material. It must collect previewable sources or stable locators for key Facts before drafting the final report.

For company/candidate ranking, valuation, liquidity, or estimate-revision discussion, a suitable financial baseline is required: yfinance, a-stock-data, Funda, TradingView, official filings/exchange data, or a documented manual equivalent. Search snippets and article summaries are not financial data.

For anti-hype mode, social/narrative origin must be captured from an original source path: OpenCLI, browser automation, social reader, or a manual original URL. If login/session/tool access is needed, ask permission and mark the channel blocked until approved.

Hard source gate

Before moving from /source-map to serious synthesis, set a hard source gate status:

• PASS: Source Registry exists; key Facts have citation ids and previewable URLs/stable locators; required market/source-specific channels are available or explicitly not material.
• FAIL: any hard source gate rule from research-tool-stack.md fails, including missing clickable original-source URLs, local files used as proof for external Facts, web-search-only evidence, missing financial baseline for company ranking, or missing OpenCLI/browser/manual narrative capture for anti-hype mode.

If the gate is FAIL:

1. Stop the workflow before /report-quality final report generation.
2. Produce a Source Gate Failure report.
3. Produce a Source Acquisition Plan with exact missing sources, preferred playbooks/tools, and user actions needed.
4. Ask for permission if local code, scripts, dependencies, API keys, browser login, or subscriptions are needed.
5. Do not generate polished report-source.md or report.html.

Before final delivery, rerun the source gate against the final HTML draft. A previous PASS is invalid if the final report contains non-clickable citations, local artifact links as evidence, ASCII diagrams instead of proper figures, or uncited data visuals.

A user may ask for a rough working note, but it must be labeled Draft — Source Gate Failed, not a final Alpha Research report.

Workflow

Use this default /alpha-research workflow:

/thesis-grill → /source-map → /chain-trace → /thesis-audit → /report-quality

Research modes:

• normal: balanced evidence-based research.
• anti-hype: stricter narrative scrutiny. In Chinese reports, explain this as “反炒作审查：专门检查热门叙事是否被市场过度放大”。

Anti-hype mode tightens the same stages:

• /source-map: trace original narrative sources.
• /thesis-audit: keep weak sources weak and emphasize crowding, reflexivity, and narrative risk.
• /report-quality: synthesize supported claims vs hype, remove promotional language, require clickable citations, teach the topic in plain Chinese, present positive/negative research views, and render Dense Calm + ECharts HTML with compact tables and ECharts SVG figures instead of ASCII diagrams.

Stage management

At the end of every stage, update run-log.md and produce a compact handoff:

## Stage Handoff

- Completed stage:
- Key decisions:
- Evidence gaps:
- Next skill:
- Handoff input:
- Stop / continue recommendation:

Keep handoffs short. Do not duplicate full stage outputs.

Internal source system

/alpha-research owns the source system. Do not route users to separate setup or provider skills.

Use internal resources:

• resources/source-profiles.md for required source mixes by research object.
• resources/source-recipes.md for acquisition sequences.
• resources/tool-install-sources.md for canonical install sources, package names, verification checks, and user approval prompts.
• resources/data-providers/*.md for source-specific guidance such as SEC EDGAR, HKEXnews, CNINFO, company IR, yfinance, a-stock-data, Funda, TradingView, OpenCLI, AkShare, Tushare, Semantic Scholar/arXiv, patents, Grok/X, social sources, China policy sources, EIA, IEA/OPEC/JODI, CME/ICE, CFTC COT, World Gold Council/LBMA, and FRED/Treasury macro sources.
• templates/tool-status.md for tool inventory.
• templates/source-gate.md for hard source gate results.

When a required tool or install source is missing, stay inside /alpha-research: produce tool-status.md / source-gate.md, ask for user approval, then continue source acquisition if approved.

Required stage order

Do not skip /source-map before serious research.

Information quality determines report quality. An Alpha Research report without a source plan, Source Registry, starter evidence pack, hard source gate PASS, and citation ids for key Facts is not acceptable.

Guardrails

• Do not present candidate recommendations as personalized allocation advice.
• Do not output buy/sell commands.
• Do not output target prices or position sizing.
• Do not promise returns.
• Do not recommend a company without source-gated evidence, clear logic, risks, counterarguments, and failure criteria.
• Do not create a trading bot or automated ordering workflow.
• Do not turn weak evidence into confident prose.
• Do not let final report quality hide evidence gaps.