sm-stock-screen

SM Stock Screen

This skill is responsible for turning "I roughly want to look at a direction" into "which 5-10 stocks should I research first."
When the user explicitly requests multiple strategies, consensus picks, ClawHub skill, or a comparison of 10 strategies, it is also responsible for incorporating external stock-picking skills into a unified comparison framework.

Mandatory process (v0.2 hard constraints)

⛔ Before any analysis output, the 6-step pre-start process of ../../core/preamble.md must be strictly executed

⛔ Before any output is completed, the 8-step post-completion process of ../../core/postamble.md must be strictly executed

Output archiving follows the ../../core/output-archive.md naming convention Output acceptance is self-checked item by item against the ../../core/acceptance.md checklist

Special notes for Stock Screen:

• Define the problem first, then screen stocks; do not directly list a bunch of names
• The screening logic must be written clearly: whether it is boominess, supply bottleneck, valuation switch, low-gain catch-up, or expectation gap
• The output is not a conclusion, but research priority
• In multi-strategy mode, define comparable criteria first, then compare results

First determine which mode it is

Mode A: Standard stock screening

Applicable to:

• "What is still missing in the AI chain"
• "Help me dig out a few key targets in MLCC"
• "Which sectors are AI-related but have risen relatively little"
• "Give me a stock-picking idea template, do not directly list stocks"

Mode B: Multi-strategy consensus comparison

Applicable to:

• "Pull all these stock-picking skills in for comparison"
• "I have 10 strategies, see what they commonly recommend"
• "Run a registry of ClawHub skills related to quantitative stock picking"
• "Which stocks are multi-strategy consensus, and which were selected by only a single style"

Reading rules

• When entering Mode B , first read references/strategy-compare-contract.md
• When entering Mode B ,mustread references/clawhub-quant-registry.md, and treat it as thedefault canonical roster of this mode
• When the user requests to refresh, add/remove, or supplement external strategies , then read references/strategy-refresh-playbook.md
• Do notautomatically switch to a local roster just because the current workspace happens to have certain local skills installed; unless the user explicitly says "switch to locally installed skills" or "include local skills in the comparison"

Five default screening methods (Mode A)

1. Bottleneck segment screening method

Suitable for: "What is still missing in the AI chain?"

• Identify segments already validated by the mainline
• Find overlooked but necessary components / materials / equipment in the supply chain
• Focus on segments where "volume rises first and profit reflects later"

2. Low-gain catch-up screening method

Suitable for: "Which sectors are AI-related but have risen relatively little?"

• First confirm what the mainline has already risen
• Find companies with the same beneficiaries but lagging gains / not heavily positioned / expectations not yet fully met
• Exclude pure left-side value traps

3. Expectation gap screening method

Suitable for: "Which names has the market not yet fully priced in?"

• The industry logic has already reached consensus
• Company-level variables have not yet been fully revised upward
• See whether there are verification points in the next 1-2 quarters

4. Industry chain diffusion screening method

Suitable for: "Where does the main theme expand from the leaders?"

• First identify the leaders and direct beneficiaries
• Then look at second-order / third-order beneficiaries
• Determine whether the diffusion runs on earnings, elasticity, or sentiment

5. Mispricing repair screening method

Suitable for: "Fundamentals are relevant, but it rose little / fell a lot for other reasons"

• First confirm whether it is truly relevant
• Then find the reason for the mispricing: short-term earnings miss, positioning, liquidity, drag from non-core businesses
• See whether the repair catalyst is clear

Multi-strategy consensus mode (Mode B)

The core of Mode B is not "piling a bunch of skill lists together", but doing three things:

1. Build a roster: First clarify which strategies to include this time, and which are only data layers / validators / routers
2. Unify the standard: Map the outputs of different strategies into a unified schema, then do raw/family-adjusted consensus
3. Explain the differences: Explain why consensus names overlap, and why divergent names are selected only by a certain style

Hard rules of Mode B

• By default the roster only recognizes references/clawhub-quant-registry.md strategies registered in; for strategies that are not registered, even if they are available in the current workspace,they cannot be automatically included
• Locally installed skills are not equal to the mode B roster; local skills can enter only under two conditions:
  1. It has a corresponding entry in the registry owner/slug / strategy card
  2. The user explicitly requests switching to local roster mode
• External skills are treated by default as strategy cards, and are not equal to actually being installed into the current workspace
• Only generator can directly vote to produce a candidate pool
• validator can only pass/veto/downgrade, and cannot create a universe on its own
• data / meta / router skills do not count as votes by default
• Approximate strategies in the same family cannot be counted as separate votes
  • For example, tvscreener and tradingview-screener belong to the same TradingView family; raw can count both, but family-adjusted can only count as 1 vote

Output goals of Mode B

• 共识票: Targets hit simultaneously by multiple independent strategy families
• 分歧票: Targets hit by only a single path, or vetoed by a validator
• 原理归因: Exactly which logic each name was selected by among trend, capital flow, valuation, macro, sentiment, and theme diffusion
• 下一步研究顺序: Which few are worth adjusting sm-thesis / sm-company-deepdive

Output format

Mode A: Standard stock selection output

# Stock Screen · {主题}

## 一句话筛选命题
- {例如“AI 训练链已经交易到光模块，但 MLCC / 高频被动件的二阶受益尚未充分定价”}

## 先筛什么，不筛什么
- 纳入：{市场 / 子行业 / 市值段 / 业务暴露}
- 排除：{纯概念、业务相关度弱、数据无法验证}

## 候选池

| 分层 | 股票 | 为什么入池 | 现在最大的疑问 | 建议下一步 |
|---|---|---|---|---|
| A 级优先看 | ... | ... | ... | 调 `sm-thesis` |
| B 级观察 | ... | ... | ... | 等数据 / 等催化 |
| C 级备选 | ... | ... | ... | 暂不深挖 |

## 这次用的筛法
- 瓶颈环节 / 低涨幅补涨 / 预期差 / 产业链扩散 / 错杀修复

## 可复用模板
1. 先问：主线已经涨了谁？
2. 再问：谁是真受益但还没涨够？
3. 再问：缺的验证数据是什么？
4. 最后分成 A/B/C 三层，不要一次全做深度

Mode B: Multi-strategy consensus output

# Strategy Compare · {主题 / 市场 / 时间窗}

## 这次比较的问题是什么
- {筛选命题}

## 本次纳入的 strategy roster
| strategy | family | role | vote_mode | 市场 | 出处 |
|---|---|---|---|---|---|

## 为什么这些 strategy 可比 / 不可比
- generators: 负责产出候选池
- validators: 负责对候选池做通过/否决
- data/meta: 不计票，只补数或编排

## 候选池并集（union pool）
| 股票 | 首次命中 strategy | 主逻辑标签 | 备注 |
|---|---|---|---|

## 共识矩阵
| 股票 | raw_hits | family_hits | validator_pass | validator_reject | 主逻辑 | 结论 |
|---|---:|---:|---:|---:|---|---|

## A级共识票
| 股票 | 为什么被多策略同时选中 | 主要分歧点 | 建议下一步 |
|---|---|---|---|

## B级交叉风格票
| 股票 | 命中哪些风格 | 为什么暂不升 A | 建议下一步 |
|---|---|---|---|

## C级单一策略票 / 分歧票
| 股票 | 只被谁选中 | 可能的偏见或限制 | 是否保留观察 |
|---|---|---|---|

## 这些重合是怎么来的
- 趋势共识 / 估值共识 / 主题扩散 / 资金流向 / 情绪催化 / 宏观门控

## 下一步研究顺序
1. {A级里最值得调 `sm-thesis` 的票}
2. {需要 `sm-consensus-watch` 补预期差的票}
3. {需要 `sm-company-deepdive` 深挖的票}

Typical trigger examples

• "What is relatively scarce in the AI chain right now?"
• "Help me dig out a few key targets in MLCC and TWS"
• "Who else is in AI-related sectors but has risen relatively little"
• "Collect the stock-picking skills on ClawHub and make a comparison"
• "I have 10 strategies here, see what they commonly recommend"
• "Give me a consensus stock framework, don't just report stocks"

Constraints

• ❌ Do not directly give a "must-rise stocks" list
• ❌ Do not force weakly related companies into the AI chain
• ❌ Do not write the stock selection candidate pool as an in-depth report
• ❌ Do not treat the data / meta / router skill as a voting machine
• ❌ Do not let duplicate strategies from the same family be weighted infinitely
• ❌ When the user has not explicitly stated it, use the local skill "conveniently available" in the current environment to replace the Claude Hub strategy in the registry
• ✅ Must be layered (A/B/C)
• ✅ Must explain this screening method or strategy roster
• ✅ Must clarify which skill to call next for deeper investigation
• ✅ When the user requests an external strategy, the source must be explicitly written (at least owner/slug or source code link)

Output archive path

themes/{theme-slug}/stock-screen/{YYYY-MM-DD}-{screen-slug}.md

Relationship with other skills

Relationship	Description
Upstream	Can first call sm-industry-map to build the industry chain framework
Parallel	Mode B can reference references/clawhub-quant-registry.md to make the strategy roster
Downstream	A-level candidates enter sm-thesis / sm-company-deepdive
Complementary	sm-consensus-watch is responsible for expectation gap verification,sm-stock-screen is responsible for finding the pool first
Audit	Registry refresh and additions are executed according to references/strategy-refresh-playbook.md Execute

Practical reminders

• When the user says "install these skills", by default understand it asinstall into the stock selection method section / registry, not immediately install all external skills
• Only when the user explicitly requests to install a certain external skill does it enter the real installation process
• By default, do the registry, comparison, consensus, and research order first; separate "method orchestration" and "real execution"
• If the user has not specifically stated otherwise, Mode B's strategy roster is based on clawhub-quant-registry.md and does not do local automatic discovery