# Investment Skills Lexicon: How great investors think

> Source: <https://github.com/finterm-ai/investment-skills>
> Published: 2026-07-22 17:04:44+00:00

**How great investors think, distilled from primary sources.**

The primary Buffett and Berkshire corpus runs 72 documents and about 950,000 words. This project extracts, analyzes, and condenses it (3,809 extracted insights and roughly 3.4 billion tokens of processing) into 65 decision cards that an agent loads 5–12 at a time.

You can always ask an agent to “analyze this like Warren Buffett.” But the goal is not to role play an investor’s voice. Rather, we want an agent to apply the underlying frameworks and mental models used by that investor to real filings, memos, and theses without searching or reading the whole corpus into context.

The idea of this project is to spend the tokens and do heavy thinking once, then reuse the insights in a context-efficient way.

This works with Claude Code, Codex, or any agent that reads the open
[Agent Skills](https://agentskills.io) standard.

**The Buffett investment framework** distills ~2,300 pages of primary Buffett and
Berkshire writing into **65 evidence-bounded decision cards** an agent can use.

It is an agent skill for **evidence-bounded financial analysis, investment-memo review,
and thesis evaluation**. It offers **three focused workflows** that surface assumptions,
counterarguments, missing evidence, and invalidation conditions.

The framework helps an agent reason about a business.
It does not issue buy, sell, hold, entry-price, position-size, or trade instructions.
The complete agent instructions are in
[SKILL.md](/finterm-ai/investment-skills/blob/main/skills/buffett-investment-framework/SKILL.md), which is self-contained once
installed.

| Stage | Contents | Scale |
|---|---|---|
| Source corpus | 48 Berkshire shareholder letters (1977–2024), Greg Abel’s 2025 transition letter, the 1957–1970 Buffett Partnership letter compilation, other Buffett and Berkshire writings, The Essays of Warren Buffett, and Buffett: The Making of an American Capitalist |
72 documents (2,300 pages, 950,000 words) |
| Extraction | Source-native insights distilled from the corpus | 3,809 insights |
| Processing | Extraction and synthesis ran across Codex Sol and Claude Fable | about 3.4 billion tokens |
| Skill lexicon | Final task-facing output | 65 cards · 8 modules · about 51 pages |

This “lexicon” is a growing collection of insights. Decades of investment thinking sit in writings, interviews, and filings that few people or even agents have time to analyze deeply. So we’re extracting these piece by piece.

The first piece is the Buffett framework. Queued next: Munger, Graham, Marks, Fisher, Lynch, Sleep and Zakaria, Grantham, Greenblatt, and Druckenmiller.

Check back here as we’re adding a new one each week!
Primary sources and framework focus for each are in the [roadmap](/finterm-ai/investment-skills/blob/main/docs/roadmap.md).

Install with the [ skills CLI](https://github.com/vercel-labs/skills) (more options
under

[Other Install Methods](#other-install-methods)):

```
npx skills add finterm-ai/investment-skills --skill buffett-investment-framework --yes
```

Then ask in your own words and attach the evidence. The skill picks the workflow and loads only the cards that fit:

```
What are the real owner earnings in these filings?
Review this investment memo and tell me which claims hold up.
Here is my thesis on this company. What would prove it wrong?
```

Most agents load the skill from its description, so naming it is optional.
In Codex, type `$`

to mention it explicitly.

| Workflow | Starting load | What it returns |
|---|---|---|
| Financial analysis | `F01` , `F02` , `F05` , `F06` , `F07` |
Filing inventory, reported-to-owner bridge, normalized segments, returns, obligations, assumptions, and blocked calculations |
| Memo review | `D01` , `D03` , `B02` , `V01` , `V05` , `V06` , `R01` , `R07` |
One dispositioned row per material claim, with counterevidence and missing evidence |
| Thesis evaluation | `D02` , `D03` , `B01` , `B06` , `V01` , `V04` , `R01` , `R07` |
Component map, mechanism and valuation tests, owner-harm paths, and invalidation conditions |

The router starts with 5–8 cards, adds only material management, allocation, financing, or specialized overlays. It caps at 12 cards in one pass to preserve context and stay focused.

| Module | Cards | Decision use |
|---|---|---|
|

`D01`

–`D09`

[Business economics](/finterm-ai/investment-skills/blob/main/skills/buffett-investment-framework/references/02-business-economics.md)`B01`

–`B07`

[Management and governance](/finterm-ai/investment-skills/blob/main/skills/buffett-investment-framework/references/03-management-governance.md)`M01`

–`M08`

[Financial reality](/finterm-ai/investment-skills/blob/main/skills/buffett-investment-framework/references/04-financial-reality.md)`F01`

–`F07`

[Valuation](/finterm-ai/investment-skills/blob/main/skills/buffett-investment-framework/references/05-valuation.md)`V01`

–`V06`

[Capital allocation](/finterm-ai/investment-skills/blob/main/skills/buffett-investment-framework/references/06-capital-allocation.md)`C01`

–`C09`

[Risk and monitoring](/finterm-ai/investment-skills/blob/main/skills/buffett-investment-framework/references/07-risk-monitoring.md)`R01`

–`R07`

[Specialized overlays](/finterm-ai/investment-skills/blob/main/skills/buffett-investment-framework/references/08-specialized-overlays.md)`S01`

–`S12`

Every card uses the same contract: decision question, guidance, use condition, analytical actions, observable output, limits, readable source basis, and abbreviated corroboration citations.

Every completed analysis reports:

- Question, horizon, scope, and exclusions.
- Evidence received and material missing inputs.
- The 5–12 cards loaded and why.
- Sourced calculations, bridges, and mechanism tests.
- Supported, challenged, or unresolved findings.
- Counterevidence and alternate mechanisms.
- Limits and blocked branches.
- Monitoring evidence and invalidation conditions.

The result ends with an analytical summary.

The framework is an **editorial synthesis of published Buffett and Berkshire writings**,
not a transcription or an attempt to imitate Buffett’s voice.
The development corpus drew primarily from Berkshire Hathaway annual letters, Buffett’s
2015 50th-anniversary essay, and “The Superinvestors of Graham-and-Doddsville.”

The material was distilled in four steps:

- Extract claims, definitions, analytical tactics, examples, and source references while preserving their source identity.
- Inventory the recurring decision questions before applying a product taxonomy.
- Reconcile overlapping questions by analytical consequence, splitting items when their evidence needs or failure conditions differ.
- Project the result into 65 consistent cards, eight modules, and three task workflows.

```
flowchart LR
    corpus["<b>Source corpus</b><br/>primary writings, letters,<br/>interviews, filings"]
    extract["<b>1 · Extract</b><br/>source-native insights,<br/>identity preserved"]
    inventory["<b>2 · Inventory</b><br/>recurring decision questions,<br/>before any taxonomy"]
    reconcile["<b>3 · Reconcile</b><br/>overlapping questions merged<br/>or split by consequence"]
    project["<b>4 · Project</b><br/>consistent cards, modules,<br/>and task workflows"]
    corroborate["<b>Corroborate</b><br/>every card re-checked against<br/>the full landed collection"]
    validate["<b>Validate</b><br/>one command: contracts,<br/>sources, routing, links"]
    use["<b>Agent use</b><br/>router loads a small card<br/>set per task, never all"]

    corpus --> extract --> inventory --> reconcile --> project
    project --> corroborate --> validate --> use

    classDef src fill:#f5f0e1,stroke:#b8a878,color:#1f2328;
    classDef step fill:#e9f2fc,stroke:#7ba7d7,color:#1f2328;
    classDef check fill:#e3f3ee,stroke:#6fb39c,color:#1f2328;
    classDef out fill:#f1ecfa,stroke:#a68fd8,color:#1f2328;
    class corpus src
    class extract,inventory,reconcile,project step
    class corroborate,validate check
    class use out
```

The published cards provide representative, not exhaustive, coverage. The process prioritized recurring decision questions rather than attempting to map every passage in the source material.

Each published card names one to three representative Buffett or Berkshire sources and the role each played: defining, supporting, implementing, illustrating, or qualifying the guidance, or specializing it for one context. These source notes explain the synthesis; they are not quotations, exhaustive literature reviews, independent corroboration, or a substitute for checking the original writing and the company evidence under analysis.

After synthesis, every card was reconciled against the project’s full landed source
collection: Berkshire letters from 1977 through 2024, the 2025 transition and farewell
letters, the partnership letters, the Owner’s Manual, the 50th-anniversary essays,
Buffett’s later public letters and comments, *The Essays of Warren Buffett* arrangement,
and the Lowenstein biography.
Each card carries abbreviated corroboration citations naming additional checked
locations where its point is stated, applied, or qualified; the
[source key](/finterm-ai/investment-skills/blob/main/skills/buffett-investment-framework/references/00-source-key.md) resolves
every abbreviation, explains the evidence character of each source, and gives a
verified retrieval location for every citation key.

This project is not affiliated with, approved by, or endorsed by Warren Buffett or Berkshire Hathaway.

**Deterministic routing:**`scripts/framework.py`

is a standard-library Python script; no model chooses the card load. The same intent and topics always produce the same cards.**A hard card cap:** 5–12 cards per pass; agent context is treated as a budget.**A one-command validator:**`validate`

checks the 65-card namespace and per-card field contracts.**A structural authority boundary:** cards return supported, challenged, or unresolved findings.

The `npx skills add`

command in [Try It](#try-it) installs into whichever agent
directories the CLI detects.
Add `-g`

to install for your user instead of the current project, `--copy`

to install an
independent snapshot rather than a symlink, and `--list`

to see what a repository offers
without installing it.
The CLI attempts every agent it detects; a per-agent failure line for an agent you do
not use (for example, PromptScript declining global installs) leaves the successful
installs intact.

To install without the CLI, copy the skill folder into the directory your agent reads. From this repository’s root:

```
# Cross-agent project install (Codex, Cursor, and pi read this path natively)
mkdir -p your-project/.agents/skills
cp -R skills/buffett-investment-framework your-project/.agents/skills/

# Claude Code
mkdir -p your-project/.claude/skills
cp -R skills/buffett-investment-framework your-project/.claude/skills/

# Codex
cp -R skills/buffett-investment-framework "${CODEX_HOME:-$HOME/.codex}/skills/"
```

Development docs live in [docs/development.md](/finterm-ai/investment-skills/blob/main/docs/development.md) and
[viz/README.md](/finterm-ai/investment-skills/blob/main/viz/README.md).

MIT. See [LICENSE](/finterm-ai/investment-skills/blob/main/LICENSE).

This project is open-source reference material, **not investment advice**. finterm.ai
makes no recommendation to buy, sell, or hold any security and accepts no liability for
investment decisions you or your agents make.