Quant Agent Skills

Production-grade quant research orchestration skills for Claude Code / Codex: concept research, single-stock research, and industry research. Built from 6+ months of daily use on a live A-share/HK/US research loop — these are not prompt collections, they are ~8,000 lines of encoded orchestration rules: multi-agent debate contracts, narrative lifecycle mapping, expectation-delta gates, closeout ordering, and three-place knowledge persistence.

What's inside

Skill	Version	Job
concept-research	1.3.2	L1 panoramic scan → L2 quick eval → L3 deep dive on market narratives; maps narrative lifecycle to best vehicles and hands off to stock research
stock-research	2.46.0	Single-stock L2+/L3 for A-share / H-share / US / A+H / ADR-HK: 6-agent evidence, market adapters, overlays, closeout contract, five-file output
industry-research	1.8.4	Top-down industry mispricing detection: Check-First, fresh public signal sweep, 5-agent debate, expectation-delta guard, five-file output

Skill content is written in Chinese (the author's working language); this README is bilingual. English translation of the core contracts is on the roadmap.

Repository layout

quant-agent-skills/
├── concept-research/      # narrative lifecycle → best vehicle
├── stock-research/        # single-stock L2+/L3 orchestration
├── industry-research/     # top-down industry mispricing detection
├── examples/              # fictional L3 five-file package (closeout demo)
├── docs/adaptation-guide.md  # glossary: which components are yours to swap
├── template/              # minimal workspace skeleton (SSoT file contracts)
└── scripts/               # publish pipeline + sanitization guard + smoke test

Why this is different

• Agent evidence contracts, not one-shot prompts. Deep research runs real sub-agents (bull/bear, financial, industry, catalyst, risk, report) with explicit fallback semantics — local review is never disguised as agent output.
• Closeout discipline. Every research package ends with machine-checked gates: verdict + trigger + watchlist + bridge + loop board + Vault sync + quality gate, in a fixed order. Warnings and pending states are never wrapped as DONE.
• Anti-look-ahead and anti-optimism rules. Expectation-delta guard treats announcements as event_occurred=true, not thesis validation; IN_LINE only triggers review, never an upgrade.
• Adaptation honesty. The skills come from a specific local stack. The glossary in docs/adaptation-guide.md tells you exactly which components are yours to substitute.

Example output

See examples/ — a complete, fictional L3 five-file package (company, numbers, and reports are invented) showing what the closeout contract produces: thesis → evidence chain → valuation → triggers → tracking. It also demonstrates the safety contract: no positions, no buy/sell actions, research conclusions routed to re_research/watch, never to trades.

Quick start

git clone https://github.com/Howard-Jerry/quant-agent-skills
cd quant-agent-skills

# Claude Code
ln -s "$PWD/concept-research"   ~/.claude/skills/
ln -s "$PWD/stock-research"     ~/.claude/skills/
ln -s "$PWD/industry-research"  ~/.claude/skills/

# Codex (or copy instead of symlink)
ln -s "$PWD/concept-research"   ~/.codex/skills/
ln -s "$PWD/stock-research"     ~/.codex/skills/
ln -s "$PWD/industry-research"  ~/.codex/skills/

If a skill with the same name already exists in your skills directory, back it up first (or copy instead of symlink). Then read the 公开版适配说明 block at the top of each SKILL.md (it links the glossary) and docs/adaptation-guide.md before your first run. The minimal workspace skeleton is in template/. To sanity-check your clone, run bash scripts/smoke-test.sh — it simulates a fresh install.

Research only. This project ships methodology, not positions: no portfolio data, no trade history, no API keys. Nothing here is financial advice.

Feedback

• Issues — methodology bugs, missing gates, or something that burned you. War stories with anonymized details are the fastest way to improve the rules.
• Discussions — how you adapted the skills to your own data stack.
• PRs — welcome for docs/, template/, and scripts/. Skill bodies are synced from the author's private SSoT via scripts/publish.sh; propose methodology changes as an Issue instead (see docs/adaptation-guide.md).

Roadmap

• v1.1: open-source the data layer (collect-data) and candidate queue (l3-candidate-queue) that the skills reference as companions.
• v1.2: English translation of the core contracts.

License

MIT. See LICENSE.

---

量化研究 Agent 三件套（开源版）

面向 Claude Code / Codex 的量化研究编排 skill：概念研究、个股研究、行业研究。 不是 prompt 拼盘——是作者在真实 A 股/港股/美股研究闭环上跑了 6 个多月、 约 8000 行踩坑换来的编排规则：多 Agent 证据契约、叙事生命周期、预期差闸、 closeout 顺序、三地沉淀。

三个 skill 的版本与职责见上方表格。安装方式见 Quick start。公开版假定你 有自己的数据层和知识库，适配点全部列在 docs/adaptation-guide.md 与 template/ 里。安装后可用 bash scripts/smoke-test.sh 自检。想看产出长什么样，直接读 examples/ 里的虚构 L3 五文件包（全部数字为编造，仅演示契约）。

反馈与协作：方法论问题、踩坑案例请开 Issue；适配经验欢迎在 Discussions 交流；docs/、template/、scripts/ 接受 PR，skill 正文按发布管线从 作者私有 SSoT 同步，方法论变更请走 Issue。

免责声明：仅供研究与教育用途，不构成投资建议；仓库不含任何实盘持仓数据。