Investment Memo Craft

Purpose

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.

Core Workflow

1. 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".

1. 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.

1. 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.

1. 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".

1. 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.

1. 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.

1. 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.

1. 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.

1. 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.

Style Standards

• 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.

Layout Standards

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.

Default Report Shape

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

1. AI研究偏见自觉

• State the information-richness rating, consensus trap, bias checklist, and AI research limitation.

1. 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.

1. 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.

1. 第三步：护城河评估

• Score moat sources, explain evidence, and state what can destroy or weaken the moat.

1. 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.

1. 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.

1. 第六步：行业与文明趋势

• Separate broad trend from investable economics and explain where the company captures value.

1. 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.

1. 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.

1. AI分析置信度 vs 投资确定性

• Separate data confidence from investment certainty.

1. 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.

Quality Bar

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?

Pairing With Other Skills

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.