agentleFS
MadkyotigerMadkyotiger/investment-os/skills/investment-research/SKILL.md

investment-research

Skill1 starsChanged 6 days agoMITView on GitHub

Routes and executes evidence-first investment research across quick market questions, Daily Scan, cross-market wrap, A-share institutional observation, triggered deep reads, named issuer studies, source audits, and reusable research systems. Use for source-backed market research, watchlists, filings, data pipelines, evidence ledgers, longitudinal analysis, portfolio-risk review, or reader briefs. Do NOT use for buy/sell/hold calls, position sizing, target prices, brokerage execution, return promises, or regulated investment advice.

8 other projects use a version of this.

Adopted by

Most recently taken up by

The file

Show its rules and text

We split this skill into the rules it gives an agent

21 rules. 0 are common, 14 are distinctive, and 7 are very specific to this project.

  • 0

    Common

    Most of the 6 edited versions we read keep these.

  • 14

    Distinctive

    This author's own choices. Worth a look.

  • 7

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Use when the user asks to:

    line 29
  2. Distinctive:

    run a Daily Scan, cross-market wrap, A-share institutional observation, deep read, issuer comparison, or evidence audit.

    line 32
  3. Very specific:

    Friend/client/external: use client-safe as an output overlay after the research route.

    line 75
  4. Very specific:

    Engineer/operator: prefer engineer-receipt; keep market interpretation out unless the failure changes evidence quality.

    line 76
  5. Distinctive:

    Several markets: use market-specific sources, then synthesize only mechanisms that really connect them.

    line 79
  6. Distinctive:

    When the user invokes an existing research workflow from chat, treat these phrases as operational commands rather than requests for shell …

    line 83
  7. Distinctive:

    Default length: one phone screen unless the decision genuinely needs more.

    line 100
  8. Distinctive:

    Keep facts, interpretation, and unknowns legible without turning the prose into a worksheet.

    line 101
  9. Distinctive:

    No internal filenames, paths, commands, adapter names, evidence-row counts, schema labels, or QA scaffolding.

    line 104
  10. Distinctive:

    Use only when requested, when the evidence is consequential, or when a reader needs to inspect trust.

    line 108
  11. Very specific:

    Every formal client-safe or otherwise forwardable report must preserve the reader/scope, as-of interval, source boundary, material facts, …

    line 131
  12. Distinctive:

    Use this fixed Chinese disclaimer for Chinese forwardable output.

    line 133
  13. Distinctive:

    search the broad universe before narrowing to a theme.

    line 146
  14. Distinctive:

    stop or narrow the claim when decision-relevant evidence cannot be read.

    line 151
  15. Very specific:

    Other markets: add dedicated exchange/regulator/issuer routes as coverage requires.

    line 160
  16. Very specific:

    Before historical reconstruction, load references/judgment-kernel-and-historical-replay.md.

    line 164
  17. Distinctive:

    Collect the smallest sufficient evidence packet: Prefer primary sources; record actual source, time, access, field semantics, and missing …

    line 170
  18. Distinctive:

    Stress test: Check source strength, stale/partial data, inference leakage, advice leakage, private residue, market-route errors, and …

    line 174
  19. Very specific:

    Verify and write back: Run the relevant code/test/readback when there are side effects.

    line 175
  20. Distinctive:

    every consequential claim is traceable to opened evidence or labelled calculation/inference.

    line 250
  21. Very specific:

    no buy/sell/hold, position, target, return promise, or execution advice leaked.

    line 254

“Common” is measured: kept by at least 6 in 10 of the 6 edited versions of this file we could read. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Investment Research

Overview

Build and operate reusable investment-information systems that collect, verify, compare, and explain market evidence without pretending to be a licensed adviser or an automatic trading brain.

The skill has two jobs:

  1. choose the right research mode and reader surface for the actual demand;
  2. enforce source, freshness, inference, privacy, and no-advice boundaries while doing the work.

It is not a generic report generator. A short current question, a Daily Scan, a source audit, a triggered deep read, and a data-pipeline repair are different jobs and must not collapse into one template.

Route Gate

Use when the user asks to:

  • explain a current market, macro, company, sector, filing, or watchlist change from sources;
  • run a Daily Scan, cross-market wrap, A-share institutional observation, deep read, issuer comparison, or evidence audit;
  • build or debug a reusable investment data/research workflow;
  • produce a reader brief, forwardable note, evidence ledger, or technical receipt;
  • evaluate investment-data or research repositories as capability modules.

Do not use to choose securities for the user, recommend buy/sell/hold, set position size, give target prices, connect brokerage execution, promise returns, or disguise advice as a score. Reframe those requests into comparative evidence, risks, assumptions, falsifiers, and research actions while leaving the decision with the user.

Pure financial-news editorial benchmarking is outside this skill. Generic repository operations belong to github-operations; implementation follows software-development-lifecycle, and reproducible failures follow systematic-debugging when available.

Universal Invariants

  1. Research, not decision. Allowed actions include opening a primary source, updating a watchlist, checking exposure, revisiting an assumption, monitoring a variable, setting an evidence trigger, or ignoring noise. They never become trade instructions.
  2. Goal before method. Name the user, decision context, market scope, time horizon, freshness requirement, and success criterion before choosing a provider, model, framework, or output length. Ask only when ambiguity changes the route; otherwise proceed.
  3. Facts before judgment. Separate reported facts, deterministic calculations, interpretation, assumptions, unknowns, and unsupported claims. Keep original values and derived figures visibly distinct.
  4. Source authority. Original regulator, exchange, issuer, filing, or official release beats summaries. A faithful readable copy may support reading when the original is blocked, but a search result, title, snippet, social post, or locator is not evidence by itself.
  5. Freshness is a gate. Exact-interval work needs an explicit as-of or interval, event time when known, publication time, access state, and source strength. Never convert “not found in this scan” into “no event occurred.”
  6. Multi-market by design. A-share coverage is additive, not the product boundary. Preserve independent source routes for Hong Kong, United States, China, and every other market in scope. Never send HK/US symbols through an A-share adapter or remove one market path because another was added.
  7. Fail closed on bad data. A successful process is not a usable observation. State requested, present, missing, stale, partial, no-event, source-error, and not-applicable coverage explicitly. Do not invent direction, amount, confidence, or reconciliation.
  8. Private and external surfaces differ. A client/friend-facing output contains no private holdings, local paths, internal repo names, system labels, or unreviewed conclusions.
  9. No capability theater. External repos, multi-agent personas, optimizer outputs, and polished reports are capability modules or question generators, not authority. Real source access and reproducible evidence decide.

Demand Router

Choose one primary route. If the request carries several needs, preserve the reader's decision path: deliver the primary surface first and attach a separate audit or engineer receipt only when it changes trust or action. Do not merge reader prose and system logs into one artifact.

RouteDefault surfaceUse whenOutput default
quick-answerreader-chatA bounded current fact, mechanism, or “what matters now?” questionUseful conclusion first; exact as-of; essential evidence, uncertainty, and original links or supplied source IDs; normally one phone screen.
daily-scanreader-briefBroad current browsing across a defined universeAll qualified material changes, without a hard maximum or minimum; no quota filling; research actions stay specific; silence is valid.
cross-market-wrapreader-briefA/H/US or another multi-market comparisonOrganize around cross-market mechanisms, not parallel news dumps; same-day co-movement is not causal proof; call a shared driver only when exposure and transmission evidence support it.
a-share-institutionalreader-plus-auditA-share watchlist or institutional-observation signalsReader brief first; ledger, coverage matrix, and source receipt remain available as the Audit surface.
triggered-deep-readreader-deep-readA current event materially changes a tracked questionCurrent delta → historical prior → changed mechanism → path to next material event → falsifier and research action.
named-deep-packresearch-packAn explicitly requested stock, pair, sector, or bounded question earns deeper workConfirm source sufficiency; align periods/definitions; show model and evidence gaps; do not make it the recurring default.
source-auditevidence-ledgerThe user asks whether a claim or source is supportableClaim-level verdict; original source and field semantics; fact/calculation/inference split; evidence strength and gaps.
system-buildengineer-receiptA reusable research pipeline, data model, collector, or agent workflow must be builtImplement and exercise the smallest coherent system; return tests, artifacts, and remaining gaps—not a mock reader report.
system-debugengineer-receiptA connector, endpoint, parser, schema, freshness gate, or run failsReproduce, diagnose, repair narrowly, execute again, and report the verified result.
boundary-refusalrefuse-and-rerouteThe user asks for a trade decision, position, target, return promise, or executionRefuse categorically; never imply that more or better data would authorize the agent to choose; offer evidence comparison or research questions instead.
blocked-sourceevidence-gapA decision-relevant source cannot be read or a required fact remains unavailableName the exact access/evidence gap, try an allowed authoritative/readable route, and bound the conclusion.
client-safeforwardable-briefExisting research must become safe to share with a friend, client, or external readerRemove private/system residue; preserve facts, caveats, and the fixed decision-boundary disclaimer; have the artifact author check the final wording.

Audience and multi-need rules

  • Personal reader: watchlist or exposure relevance may be used when the user supplied it; never infer hidden holdings.
  • Friend/client/external: use client-safe as an output overlay after the research route; preserve no-advice language and source limits.
  • Engineer/operator: prefer engineer-receipt; keep market interpretation out unless the failure changes evidence quality.
  • “完整 / full”: raise the evidence and verification bar. Do not dump every log, path, source row, or hidden worksheet into chat.
  • Reader + audit: publish the brief first. Put detailed evidence in a separate artifact or compact receipt.
  • Several markets: use market-specific sources, then synthesize only mechanisms that really connect them. Missing one market is a coverage gap, not permission to substitute another.

Natural-language trigger contract

When the user invokes an existing research workflow from chat, treat these phrases as operational commands rather than requests for shell instructions:

  • 跑投研简报 / 更新投研简报 / 今天的投研简报 → locate the active project owner, run its current verified pipeline when one exists, read back the generated outputs and quality result, then return the phone-screen brief first with a compact audit receipt. If no runnable pipeline exists, state that boundary instead of hand-writing a summary and calling it pipeline output.
  • 只刷新投研读者版 / 只重生成简报 → rerender the reader brief and detailed pack from the current verified ledger/queue without rerunning source collection. If the inputs are stale, label the stale boundary rather than silently refreshing the date.
  • 看详细版 / 展开详细版 → read, summarize, or return the existing detailed pack. Do not start new collection unless a named freshness or evidence gap blocks the requested readback.
  • 完整报一遍 / 完整跑一遍 / 不足的信息先补齐 → read the live project/handoff/README, run the current verified pipeline, read back the brief, detailed pack, and quality scans, then fill material gaps with fresh authoritative sources. Lead with the research judgment and keep logs in a compact receipt.
  • 跑全市场简报 → run only when a source-universe collector has real coverage for the requested markets. Otherwise name the unsupported markets and return blocked-source; never imply full-market coverage from a narrower live pipeline.

For ordinary daily triggers, paste the reader brief first. For 完整 or gap-filling work, follow it with pipeline/test status, evidence and quality status, and exact remaining source gaps. Paths are receipts, not the product.

For detailed reader style, source-universe selection, and two-tier delivery, load references/investment-output-style-and-source-universe.md.

Output Contract

Reader surface

  • First useful line: judgment, current state, or decision-relevant change—not process narration.
  • Default length: one phone screen unless the decision genuinely needs more.
  • Keep facts, interpretation, and unknowns legible without turning the prose into a worksheet.
  • Cite only sources that carry the conclusion. Preserve original URLs where possible. If the user supplies an unlinked packet, cite its source IDs when available; otherwise label the material unverified rather than presenting it as sourced current fact.
  • Natural Chinese or the user's requested language; technical market terms may remain English when they are the normal term.
  • No internal filenames, paths, commands, adapter names, evidence-row counts, schema labels, or QA scaffolding.

Audit surface

Use only when requested, when the evidence is consequential, or when a reader needs to inspect trust. It may contain:

  • source/claim ledger and original/readable URLs;
  • coverage, freshness, field semantics, calculations, and reconciliation status;
  • pipeline/test result, source receipt, and exact remaining gaps;
  • files or IDs needed for handoff.

The Audit surface does not replace the reader product. For A-share institutional observation, deliver the brief first, then the audit artifact or a compact receipt.

Mode-specific minimums

  • quick-answer: conclusion, as-of, 1–3 decisive facts, uncertainty, and traceable source links or supplied source IDs.
  • daily-scan: all qualified material changes, without a hard maximum or minimum; what deserves attention and the next decisive research variable. No repeated-theme quota.
  • cross-market-wrap: candidate mechanism, market-specific confirmation or contradiction, explicit non-causality for same-day co-movement, coverage boundary, next variable. When exposure and transmission evidence are missing, use “candidate mechanism,” “possible mapping,” or “co-movement”; do not describe the event as driving, catalyzing, causing, or leading the market move.
  • a-share-institutional: observed signal, why it may matter, inference limit, next confirmation; audit stays optional.
  • triggered-deep-read: genuinely new delta, prior state, next-event path, observable falsifier.
  • named-deep-pack: bounded question, reconciled baseline, scenario/model limits, unanswered variables.
  • source-audit: claim verdict, authoritative evidence, derivation/inference status, missing proof.
  • system-build / system-debug: verified artifact or repair, commands/tests, exact failure or residual risk.
  • boundary-refusal: refusal based on authority boundary, never on missing data; more evidence can improve comparison but cannot delegate the trade decision to the agent.

Forwardable output standard

Every formal client-safe or otherwise forwardable report must preserve the reader/scope, as-of interval, source boundary, material facts, interpretation, unknowns, research actions, and decision boundary. Do not remove caveats merely to make the note cleaner.

Use this fixed Chinese disclaimer for Chinese forwardable output; translate it without weakening for another language:

本报告只用于信息整理、研究讨论和风险提示,不构成投资建议、交易建议或收益承诺。市场有风险,任何交易决策应由使用者基于自己的风险承受能力、资金期限、税务/法律情况和独立判断作出。报告中的数据可能延迟、缺失或有误,请在行动前自行核验。

The artifact author checks formal, reusable, or forwardable prose after facts and structure are locked. Never add vividness, certainty, examples, or claims that the evidence did not earn.

Evidence and Source Gate

Current-intelligence gate

For an exact interval:

  1. resolve the timezone and freshness cutoff;
  2. search the broad universe before narrowing to a theme;
  3. open the original source or a faithful readable route;
  4. record publication time and underlying event time where known;
  5. deduplicate by event, not headline;
  6. separate included, excluded, not found in this scan, blocked-source, and source-error states;
  7. stop or narrow the claim when decision-relevant evidence cannot be read.

A source audit should retain title, publisher, timestamp, original URL, readable URL if different, access state, evidence strength, and inclusion/exclusion reason. Locator-only results stay locators.

Market routes

  • US/HK recurring market layer: configured watchlist prices may use yfinance; Stooq is a second-source check only when a live probe returns usable same-date rows. Use SEC/FRED, issuer IR, HKEX, regulators, and original releases as authority for consequential claims.
  • Optional global depth: OpenBB only when it routes to a genuinely independent provider; FinanceToolkit for normalized statements/ratios when earned; edgartools for complex filing parsing.
  • A-share branch: AKShare is the keyless adapter; Tushare is optional and credentialed. PBOC/NBS/CNINFO/SSE/SZSE and issuer disclosures remain authoritative. Do not count an absent Tushare token as a gap.
  • Other markets: add dedicated exchange/regulator/issuer routes as coverage requires. Their absence must be reported as a gap, never as evidence that the market is out of scope.

For A-share operations: route stock, ETF, and index explicitly; never send a stock to an ETF endpoint. Use a bounded Eastmoney → Sina → Tencent stock-history fallback and retain the actual source. Stale data is not usable. no_event is not source_error. Institutional-flow language is inference, not an observed actor decision. Missing or invalid LHB amounts cannot become net-buy/net-sell direction. Beijing margin data is not_applicable when the required exchange route does not exist; do not substitute SSE/SZSE.

Before historical reconstruction, load references/judgment-kernel-and-historical-replay.md; before filing retrieval/claim extraction, load references/source-bounded-filing-claim-cards.md; for an explicit COT positioning question, load the on-demand CFTC section in references/operational-pitfalls-and-verification.md. These are research procedures, not automatic collectors or new execution authority.

Core Workflow

  1. Frame the real demand. Identify route, audience, market/source scope, interval or next-event horizon, and failure cost.
  2. Check prerequisites. Resolve accessible sources, credentials, current project/source authority, and whether the request is reader-facing, audit-facing, or engineering work.
  3. Collect the smallest sufficient evidence packet. Prefer primary sources; record actual source, time, access, field semantics, and missing coverage.
  4. Reconcile before interpretation. Align periods, currencies, definitions, issuer calendars, and deterministic calculations. Do not compare unlike numbers because labels look similar.
  5. Form the judgment. State what changed, why the evidence changes—or does not change—the question, alternatives, and what would prove the read wrong.
  6. Deliver the selected surface. Reader first by default; audit or engineer receipt separately.
  7. Stress test. Check source strength, stale/partial data, inference leakage, advice leakage, private residue, market-route errors, and whether the next action stays at the research layer.
  8. Verify and write back. Run the relevant code/test/readback when there are side effects. Update only the owning project/state/evidence surface; temporary progress stays out of durable knowledge.

For building or adopting a system, load references/system-build-and-adoption-runbook.md. For a source-backed runnable mainline, also load references/source-backed-mainline-spike.md. For failures, reproduce before changing method and use the live fallback/repair contract rather than inventing success.

Reference Router

Load only the branch the selected demand needs. Every listed asset is part of the retained capability.

An installed copy may add references/local-deployment.md for routes that exist only in that deployment (private modules, sibling skills, promotion gates). Read it once when present. The public package does not ship it, and it must not change any rule in this file.

Reader products and output

  • Product A/B/on-demand selection → references/daily-scan-vs-triggered-rolling-deep-read.md
  • Scheduled agent-researched daily brief (history and open questions, first-hand official captures, packet validation, collection report) → docs/RADAR.md in the investment-os repository; commands under investment-os radar --help
  • Real Product B execution → references/triggered-rolling-deep-read-runbook.md
  • Daily versus named deep pack → references/daily-scan-vs-stock-deep-pack.md
  • Reader style, source universe, two-tier output → references/investment-output-style-and-source-universe.md
  • Technical green versus reader utility → references/product-utility-acceptance-gate.md
  • Personal-investor relevance layer → references/cxo-personal-intelligence-layer.md
  • Current hard-source change tracking → references/cxo-hard-source-change-tracking-layer.md
  • Live current-source collection → references/live-hard-source-cxo-collector.md
  • House view / multi-signal synthesis → references/cxo-house-view-and-viewpoint-synthesis.md

Evidence, issuer, and thesis work

  • Atomic filing/earnings claims → references/source-bounded-filing-claim-cards.md
  • Company baseline and capital linkage → references/company-baseline-reconciliation-and-capital-linkage.md
  • Forecast horizon / fiscal alignment → references/forecast-horizon-and-fiscal-period-alignment.md
  • Issuer IR/SEC corpus → references/issuer-ir-sec-earnings-corpus-recon.md
  • NVIDIA IR/SEC worked pattern → references/nvidia-ir-sec-source-recon-pattern.md
  • Korean/KRX issuer corpus → references/korea-issuer-ir-earnings-corpus-recon.md
  • Longitudinal research memory → references/longitudinal-research-memory-system.md
  • Historical replay / judgment kernel → references/judgment-kernel-and-historical-replay.md
  • HOLD to event-driven falsifier state → references/model-gate-hold-to-falsifier-dashboard.md
  • Hyperscaler capex to economics → references/hyperscaler-capex-to-economics-conversion-bridge.md
  • Expert signal as question fuel → references/expert-signal-crosswalk.md
  • Cross-node thesis delta audit → references/cross-node-thesis-delta-audit.md

Anchor-network research

  • Source-collection contract → references/anchor-network-source-collection-contract.md
  • Network method → references/anchor-company-network-research.md
  • Reusable assets → templates/anchor-network-comparison.md, templates/company-history-dossier.md, and templates/thesis-delta-card.md

System design and external capability

  • Full build/adoption workflow → references/system-build-and-adoption-runbook.md
  • Runnable source-backed mainline → references/source-backed-mainline-spike.md
  • Repository landscape → references/github-investment-advisor-landscape.md
  • External methodology repo audit → references/external-investment-methodology-repo-audit.md
  • Historical US/China pilot calibration → references/us-china-unified-pilot.md; treat it as a worked branch, not the market boundary.

Operational depth

  • Full failure catalogue and branch verification → references/operational-pitfalls-and-verification.md

Common Pitfalls

  • generating one generic “investment report” regardless of the user's demand;
  • exposing an engineer receipt to a normal reader, or hiding audit evidence inside polished prose;
  • treating a blocked source, missing row, stale price, or optional unconfigured provider as the same failure;
  • presenting source count, agent debate, optimizer precision, or report length as quality;
  • letting one market adapter become the product boundary;
  • turning a research action into a trade instruction;
  • making a refusal conditional on missing data, which implies that enough data would authorize the agent to choose a trade;
  • repeating a fixed deep question without a current delta;
  • calling a technically green pipeline a useful reader product before human evaluation.

Load references/operational-pitfalls-and-verification.md for branch-specific traps before material system, source, model, A-share, client-facing, or scheduled-delivery work.

Done State

Done means:

  • the demand route, audience, market scope, time horizon, and output surface are explicit;
  • every consequential claim is traceable to opened evidence or labelled calculation/inference;
  • freshness, coverage, missing/partial/no-event/source-error states are truthful;
  • multi-market routes remain independent and A-share work did not narrow the system;
  • reader output is useful without system residue; audit/engineer proof is available when consequence requires it;
  • no buy/sell/hold, position, target, return promise, or execution advice leaked;
  • code or live writes were actually exercised and read back;
  • formal/forwardable prose passed final-language QA;
  • the full branch checklist in references/operational-pitfalls-and-verification.md was used when the work was material.

Routing Eval

The executable fixture set is tests/fixtures/investment_research_routing.yaml. Representative routes:

  • “美债收益率这周发生了什么?只告诉我现在值得知道的。” → quick-answer / reader-chat.
  • “跑今天的投资 Daily Scan。” → daily-scan / reader-brief.
  • “把 A 股、港股和美股放在一起看。” → cross-market-wrap; preserve separate market evidence.
  • “看这五只 A 股有没有机构行为信号,普通读者版先给我。” → a-share-institutional; brief first, audit optional.
  • “NVDA 财报后原来的判断哪里变了?” → triggered-deep-read.
  • “完整比较 Micron 和 SK hynix 的 HBM 兑现路径。” → named-deep-pack.
  • “核对‘主力资金明显净流入’能不能成立。” → source-audit.
  • “搭一个可复用的多市场投研管线,并实际跑通。” → system-build.
  • “AKShare 的 Eastmoney 历史行情突然失败,排查。” → system-debug.
  • “这五只股票到底买哪只?” → boundary-refusal; reroute to evidence comparison.
  • “公告打不开,直接说今天没有重要事件。” → blocked-source; retrieval failure cannot become a negative fact.
  • “改成可以转发给朋友和客户的版本。” → client-safe; remove private/system residue and retain evidence limits.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.