investment-memo-craft

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

Investment Memo Craft

投资备忘录撰写指南

Purpose

目的

Turn investment research into a decision-ready Codex research report. Keep the data discipline of the underlying research skill, but make the output easier for an investor to use: concrete business mechanics, sharp inverse thinking, explicit opportunity cost, action thresholds, and calm Markdown typography.

Use this as a writing and judgment overlay. It does not replace financial-data rules, primary-source checks, valuation tools, or report audit tooling.

For long-form AI Berkshire outputs, title the artifact as a "research report" by default. Use "investment memo" only when the user explicitly asks for a memo format.

This is a Codex-only hand-written skill kept under

codex-skills/

for simple installation. Do not add a same-named

skills/investment-memo-craft.md

source unless intentionally adopting this workflow for Claude Code too.

将投资研究转化为可供直接决策的Codex研究报告。保留底层研究技能的数据严谨性，同时让输出更便于投资者使用：清晰的商业逻辑、敏锐的逆向思维、明确的机会成本、行动阈值，以及简洁的Markdown排版。

将此作为写作与判断的模板，它不会替代财务数据规则、原始资料核查、估值工具或报告审计工具。

对于伯克希尔AI的长篇输出，默认将成果命名为「研究报告」。仅当用户明确要求备忘录格式时，才使用「投资备忘录」命名。

这是一个仅适用于Codex的手写技能，存放在

codex-skills/

目录下以便简单安装。除非有意将此工作流也应用于Claude Code，否则请勿添加同名的

skills/investment-memo-craft.md

源文件。

Core Workflow

核心工作流程

Open with context; reserve the full decision for after the evidence.
- In the first screen, state the research date, price, market cap, valuation, and a short thesis.
- Do not front-load the full buy/hold/sell table unless the user explicitly asks for an executive memo.
- Put the detailed recommendation, investor-specific actions, and price bands near the end, after business quality, risk, and valuation have been argued.
- Separate "good business" from "good investment at this price".
Build the operating map before the philosophy.
- Include revenue structure, segment economics, unit drivers, and 3-5 year trends early.
- For asset-heavy businesses, show the key assets individually when they explain the moat.
- Explain the pricing mechanism, customer lock-in, cost structure, and reinvestment needs.
Compress business essence into one memorable sentence.
- Prefer a sentence that describes who pays, why they pay, what is scarce, and what repeats.
- Avoid generic labels such as "industry leader" unless followed by the mechanism that makes leadership durable.
Make the moat falsifiable.
- Score or table the moat by source: brand/pricing power, switching cost, network effect, scale, cost advantage, regulation, resource scarcity, technology.
- Explain whether the moat widened or narrowed over the last 5 years.
- Ask what can destroy the moat, even if the answer is "not competitors, but regulation/weather/price paid".
Do real inverse thinking.
- Include failure paths with probability, impact, and observable indicators.
- Write the strongest bear case in language a smart short seller or non-buyer would actually use.
- Explicitly identify the most likely analytical mistake.
Evaluate management through capital allocation.
- Replace vague praise with decision history: acquisitions, divestitures, buybacks, dividends, leverage, reinvestment, strategic pivots.
- Judge incentives: insider ownership, controlling shareholder behavior, compensation, related-party transactions, and shareholder return policy.
- Ask whether the business depends on a person or on a system.
Connect industry trend to value capture.
- Distinguish civilization-level trend from investable company-level economics.
- Describe where the company sits in the value chain and who captures the profit pool.
- Identify whether TAM growth, pricing, utilization, or capital intensity is the real driver.
Convert valuation into action.
- Show current multiples, reverse DCF intuition, scenario valuation, historical comparison, and comparable companies when relevant.
- Include dividends or capital returns in expected return when they matter.
- Provide price bands, add signals, trim/sell signals, and what would change the thesis.
Close with a decision memo.
- Include a summary table by business quality, moat, management, risk, trend, and valuation.
- Give distinct advice for empty-handed investors and existing holders.
- Include the action table here, not at the top, for long-form research reports.
- End by separating AI analysis confidence from actual investment certainty.

以背景信息开篇；将完整决策放在证据之后。
- 在开篇部分，说明研究日期、股价、市值、估值以及简短的核心论点。
- 除非用户明确要求执行备忘录，否则不要在开头放置完整的买入/持有/卖出表格。
- 将详细建议、针对特定投资者的行动方案以及价格区间放在接近末尾的位置，在论证完业务质量、风险与估值之后。
- 区分「好生意」与「当前价格下的好投资」。
先构建业务运营框架，再阐述投资理念。
- 尽早包含收入结构、细分业务经济性、单位驱动因素以及3-5年趋势。
- 对于重资产企业，当关键资产能解释护城河时，单独列出这些资产。
- 解释定价机制、客户锁定、成本结构以及再投资需求。
用一句易记的话提炼业务本质。
- 优先选择描述付费方、付费原因、稀缺性以及重复盈利点的句子。
- 避免使用「行业领导者」这类通用标签，除非后续说明支撑领导力持久性的机制。
让护城河具备可证伪性。
- 按来源对护城河打分或制表：品牌/定价权、转换成本、网络效应、规模、成本优势、监管、资源稀缺性、技术。
- 说明过去5年护城河是拓宽还是收窄。
- 探讨可能摧毁护城河的因素，哪怕答案是「不是竞争对手，而是监管/天气/收购价格」。
进行真正的逆向思考。
- 包含带有概率、影响程度以及可观察指标的失败路径。
- 用聪明的卖空者或潜在买家会使用的语言撰写最有力的看空论点。
- 明确指出最可能出现的分析错误。
通过资本配置评估管理层。
- 用决策历史替代模糊的赞美：收购、剥离、股票回购、分红、杠杆、再投资、战略转型。
- 判断激励机制：内部人持股、控股股东行为、薪酬、关联交易以及股东回报政策。
- 探讨业务是依赖个人还是依赖体系。
将行业趋势与价值捕获关联。
- 区分文明级趋势与可投资的公司级经济性。
- 描述公司在价值链中的位置以及谁占据利润池。
- 明确TAM增长、定价、利用率或资本密集度哪个是真正的驱动因素。
将估值转化为行动。
- 展示当前估值倍数、反向DCF逻辑、情景估值、历史对比以及相关可比公司（如适用）。
- 当股息或资本回报重要时，将其纳入预期回报。
- 提供价格区间、加仓信号、减仓/卖出信号，以及会改变核心论点的因素。
以决策备忘录收尾。
- 包含按业务质量、护城河、管理层、风险、趋势以及估值分类的汇总表。
- 为未持仓投资者和现有持仓者提供不同建议。
- 对于长篇研究报告，在此处放置行动表格，而非开头。
- 结尾区分AI分析置信度与实际投资确定性。

Style Standards

风格标准

Prefer concrete numbers and mechanisms over adjectives.
Use tables when they reduce cognitive load: assets, segments, failure paths, management decisions, scenario valuations, action bands.
Write in clear investor prose. A good memo should be understandable after one read and useful after one month.
Keep memorable formulations, but never let rhetoric outrun evidence.
Avoid hiding behind vague labels such as "wait and see" without specifying the price or event that would change the recommendation.

优先使用具体数字与机制，而非形容词。
当表格能降低认知负荷时使用表格：资产、细分业务、失败路径、管理层决策、情景估值、行动区间。
用清晰的投资者风格写作。一份优质备忘录应做到一读就懂，一个月后仍有参考价值。
保留易记的表述，但绝不让修辞超过证据支撑。
避免用「观望」这类模糊表述，除非明确说明会改变建议的价格或事件。

Layout Standards

布局标准

For long-form research reports, prefer a calm stepped layout:

Use a simple title:
```
公司名（ticker）研究报告
```
. Avoid adding "四大师综合" or "投资备忘录" to the title unless the user asks for that framing.
Use dated filenames for reports:
```
公司名研究报告-YYYYMMDD.md
```
.
Start with one compact metadata block: research date, price, market cap, key multiples, and a one-sentence thesis.
Use horizontal separators between major sections.
Use Chinese step headings for readability, for example "第一步：核心数据总览", "第二步：生意本质分析", and "第八步：最终决策与行动清单".
Keep section titles short and concrete; avoid dense numbering such as "2.3.1" unless the document is technical.
Use quote blocks for master-style questions, not inline bold paragraphs.
Treat GitHub Markdown as the typography system: use heading levels, tables, quote blocks, and bold text; do not add HTML/CSS font styling unless the user explicitly asks for a non-GitHub artifact.
Use bold sparingly as a reading guide: metadata labels, one-sentence conclusion labels, key phrases, total/current-company rows, latest-year values, scenario target prices, action rows, and audit verdicts.
Keep ordinary facts in normal weight. Do not bold full tables or every important-looking number; over-emphasis makes long research feel noisy.
Use explicit
```
+
```
and
```
-
```
signs for growth rates and return ranges so positive/negative movement can be scanned without rereading the sentence.
Put checklists under "AI research bias awareness" when the company is information-rich or consensus-heavy.
Keep audit and tool details light at the end. Do not expose command lines unless the user asks for reproducibility commands.
If a prior report has a layout the user likes, preserve its reading rhythm while keeping only data that passes the current validation standard.

对于长篇研究报告，优先采用简洁的阶梯式布局：

使用简单标题：
```
公司名（股票代码）研究报告
```
。除非用户要求，否则请勿在标题中添加「四大师综合」或「投资备忘录」。
报告文件名包含日期：
```
公司名研究报告-YYYYMMDD.md
```
。
开篇放置一个紧凑的元数据块：研究日期、股价、市值、关键估值倍数以及一句核心论点。
主要章节之间使用水平分隔线。
为提升可读性，使用中文步骤标题，例如「第一步：核心数据总览」、「第二步：生意本质分析」以及「第八步：最终决策与行动清单」。
章节标题简短具体；避免使用「2.3.1」这类复杂编号，除非文档具有技术性。
使用引用块呈现大师风格的问题，而非加粗段落。
将GitHub Markdown作为排版体系：使用标题层级、表格、引用块和加粗文本；除非用户明确要求非GitHub格式文档，否则不要添加HTML/CSS字体样式。
谨慎使用加粗作为阅读指引：元数据标签、一句式结论标签、关键短语、总计/当前公司行、最新年度数值、情景目标价、行动行以及审计结论。
普通事实使用常规字体。不要给整个表格或每个看似重要的数字加粗；过度强调会让长篇研究显得杂乱。
增长率与回报范围使用明确的
```
+
```
和
```
-
```
符号，以便无需重读句子就能快速识别正负变动。
当公司信息量大或存在共识陷阱时，在「AI研究偏见自觉」下放置检查清单。
将审计与工具细节放在末尾且保持简洁。除非用户要求可复现命令，否则不要展示命令行内容。
如果用户喜欢之前某份报告的布局，保留其阅读节奏，同时只保留符合当前验证标准的数据。

Default Report Shape

默认报告结构

For AI Berkshire company reports, use this order unless the user asks otherwise:

```
AI研究偏见自觉
```
- State the information-richness rating, consensus trap, bias checklist, and AI research limitation.
```
第一步：核心数据总览
```
- Show segment revenue, key operating assets or units, 3-5 year financial trend, and cross-source validation.
```
第二步：生意本质分析
```
- Define the business in one sentence, map revenue/cost/customer/asset life/growth drivers, and explain the real profit variables.
```
第三步：护城河评估
```
- Score moat sources, explain evidence, and state what can destroy or weaken the moat.
```
第四步：逆向思考与风险清单
```
- Put the bear case in serious language. Include failure paths, probability, impact, and observable warning indicators.
```
第五步：管理层评估
```
- Judge management through capital allocation, governance, incentives, dividends/buybacks, leverage, and whether the business is system-driven.
```
第六步：行业与文明趋势
```
- Separate broad trend from investable economics and explain where the company captures value.
```
第七步：估值与安全边际
```
- Show current valuation, reverse-DCF intuition, scenario valuation, comparable companies if useful, and explicit price bands.
```
第八步：最终决策与行动清单
```
- Put the full decision here, not at the top: summary table, advice for empty-handed investors, advice for holders, add/sell triggers, and master-style comments if useful.
```
AI分析置信度 vs 投资确定性
```
- Separate data confidence from investment certainty.
```
数据来源与审计记录
```
- List key sources and concise audit results. Keep command lines out of the report unless explicitly requested.

对于伯克希尔AI的公司报告，除非用户另有要求，否则使用以下顺序：

```
AI研究偏见自觉
```
- 说明信息丰富度评级、共识陷阱、偏见检查清单以及AI研究局限性。
```
第一步：核心数据总览
```
- 展示细分业务收入、关键运营资产或业务单元、3-5年财务趋势以及多源验证结果。
```
第二步：生意本质分析
```
- 用一句话定义业务，梳理收入/成本/客户/资产生命周期/增长驱动因素，并解释真正的利润变量。
```
第三步：护城河评估
```
- 对护城河来源打分，说明证据，并阐述可能摧毁或削弱护城河的因素。
```
第四步：逆向思考与风险清单
```
- 用严谨的语言呈现看空论点。包含失败路径、概率、影响程度以及可观察的预警指标。
```
第五步：管理层评估
```
- 通过资本配置、治理、激励机制、股息/回购、杠杆以及业务是否由体系驱动来评判管理层。
```
第六步：行业与文明趋势
```
- 区分广泛趋势与可投资经济性，并解释公司如何捕获价值。
```
第七步：估值与安全边际
```
- 展示当前估值、反向DCF逻辑、情景估值、可比公司（如有用）以及明确的价格区间。
```
第八步：最终决策与行动清单
```
- 将完整决策放在此处，而非开头：汇总表、给未持仓投资者的建议、给持仓者的建议、加仓/卖出触发条件，以及如有必要的大师风格评论。
```
AI分析置信度 vs 投资确定性
```
- 区分数据置信度与投资确定性。
```
数据来源与审计记录
```
- 列出关键来源与简洁的审计结果。除非明确要求，否则不要在报告中展示命令行内容。

Quality Bar

质量标准

A strong memo should answer these questions without forcing the reader to infer:

What exactly does this company sell, to whom, and why does money repeat?
What are the 2-3 variables that actually move profit?
Why might smart people refuse to buy?
What is already priced in?
What return is plausible under bull/base/bear cases, including dividends if relevant?
What should an empty-handed investor do?
What should a holder do?
What evidence would make the thesis wrong?

一份优质备忘录应无需读者自行推断就能回答以下问题：

这家公司具体卖什么、卖给谁，以及利润为何能持续？
真正影响利润的2-3个变量是什么？
为什么聪明人可能拒绝买入？
当前股价已经反映了哪些信息？
在牛市/中性/熊市情景下，合理的回报是多少（如相关，包含股息）？
未持仓投资者应该怎么做？
持仓者应该怎么做？
什么证据会证明核心论点错误？

Pairing With Other Skills

与其他技能搭配使用

When the task requires fresh company research, first use the relevant data/research skill and its validation requirements. Then use this skill to rewrite or structure the output as a memo.

For AI Berkshire work, pair especially with:

```
financial-data
```
for source hierarchy and cross-source validation.
```
investment-research
```
for the Buffett/Munger/Duan/Li Lu framework.
```
management-deep-dive
```
when management quality is the core uncertainty.
```
report_audit.py
```
before treating a report as publishable.

当任务需要全新的公司研究时，先使用相关的数据/研究技能及其验证要求。然后使用本技能将输出改写或整理为备忘录格式。

对于伯克希尔AI的工作，尤其建议搭配以下技能：

```
financial-data
```
：用于来源层级与多源验证。
```
investment-research
```
：用于巴菲特/芒格/段永平/李录框架。
```
management-deep-dive
```
：当管理层质量是核心不确定性时使用。
```
report_audit.py
```
：在将报告视为可发布前使用。