buffett-skills packages an investment checklist as a Claude Code skill

The skill fires on investment questions, not on the name

The repository is a collection of Claude Code skills built around Warren Buffett's investment framework, and it currently contains one, named `buffett`. The activation rule is the interesting part. The README lists the situations that trigger it automatically: analyzing any stock or company, evaluating an investment opportunity, reading financial reports, annual letters or shareholder letters, judging a company's moat or competitive advantage, assessing management quality and integrity, making buy, hold or sell decisions, discussing value investing concepts such as compounding, intrinsic value, margin of safety, circle of competence and Mr. Market, analysing specific industries from insurance and banking to tech, handling capital allocation, buybacks or dividends, and assessing market sentiment or macro risk. The README makes the point explicitly that you do not need to mention Buffett, since any investment analysis or business quality question triggers the skill.

Progressive disclosure keeps the context small

The mechanism for that activation is a layout rather than a prompt trick. The skill uses progressive disclosure, where `SKILL.md` is always loaded and the reference files are read on demand based on the analysis type. The folder holds a `SKILL.md` and eight files under `references/`, numbered so that the reading order is legible: `01-thinking-frameworks.md`, `02-investment-philosophy.md`, `03-business-moat.md`, `04-management-governance.md`, `05-financial-metrics.md`, `06-valuation-capital.md`, `07-risk-behavior.md` and `08-industry-playbooks.md`. Each file has a stated scope. The first covers circle of competence, inversion, Mr. Market, long term thinking and Munger's mental models, the second intrinsic value, compounding, undervaluation, concentration and the EMH rebuttal, the third the five moat types, franchise against commodity business and economic goodwill, the fourth a three dimensional management assessment, institutional imperative and corporate culture, the fifth owner earnings, ROIC and ROE, cash conversion rate and look-through earnings, the sixth three valuation methods, margin of safety and five capital allocation paths, the seventh when to sell, value traps, leverage, inflation, derivatives and behavioural biases, and the eighth industry playbooks for insurance, banking, consumer, media, energy, railroads and tech plus cautionary tales.

Three dispatch paths decide which files get read

The dispatch table in the README is what keeps the token cost down. Path A, quick screen, is for the question of whether something is worth analysing at all, and it reads no reference files because it uses an eight question checklist. Path B, deep analysis, is a full company evaluation, and it reads the files in the order 03, 04, 05 and 06, adding 08 for the industry and 07 for risk as needed. Path C, topic question, is a specific concept or decision, and it loads the relevant reference file directly. The mind map in the README names the three core modules this produces, a moat scanner with cost advantage, switching costs, network effects, intangible assets and efficient scale, a management mirror comparing promises against delivery, capital allocation history and pay to shareholder alignment, and a margin of safety calculator with optimistic, base and pessimistic cases producing an intrinsic value range and a suggested entry price. A well formed question therefore costs a checklist, and only a full analysis pays for the file reads.

Every answer has to fill the same template

The output format is where the discipline comes from, and the README says every analysis follows a structured template with all sections required. The conclusion comes first, a buy, don't buy, watch, hold or sell call with a one sentence rationale. Then the circle of competence verdict, in circle, out of circle or boundary, and the three to five key assumptions the decision depends on. Business quality follows with the moat type, its strength and trend, plus a management assessment, then a financial snapshot with ROIC, cash conversion rate and an owner earnings estimate, then valuation with an intrinsic value range, a margin of safety percentage and a suggested entry price. The sell criteria check is required for hold and sell questions, with four criteria checked individually. The closing sections are the top three key risks, the monitoring indicators with quarterly checks and sell triggers, and a final verdict written in Buffett's voice. The reason this is worth knowing is that a fixed structure is auditable. Two analyses of different companies can be compared field by field, and a missing section is visible rather than hidden in prose.

Installing it takes a clone and a copy

The installation is deliberately manual, with two commands. You clone the repository and copy the `skills/buffett` folder into your project:

git clone https://github.com/agi-now/buffett-skills /tmp/buffett-skills
cp -r /tmp/buffett-skills/skills/buffett your-project/.claude/skills/buffett

The skill has to end up at `.claude/skills/buffett/SKILL.md` relative to the project root, and the README is explicit that Claude Code auto-discovers skills placed there, with no registration required. That is the whole setup, and the consequence is that the skill is a file in your repository, which means it is versioned with your code, reviewable in a pull request, and copied into a new project by hand when you want it there. The repository also carries a Chinese version of the README at `README.zh.md`, a demo video at `bft.mp4`, and the `skills/` directory at the root.

The reported evaluation and what it cost

The README includes an eval table, and it is worth reading both halves. The skill was benchmarked across three test cases, an Apple stock analysis, a banking framework and a Moutai hold or sell decision, and the table compares the with-skill and without-skill runs. The pass rate is reported as 100 percent, 15 of 15, against 66.7 percent, 10 of 15, a delta of plus 33 percent. The cost is on the next two rows. Average tokens rise from 12,524 to 43,353, and average time from 95.2 seconds to 154.4 seconds. The README attributes the primary value to systematic reference file reading, at 11 or more tool calls against 2, and to strict output format adherence, and calls the token and time cost the tradeoff for analytical depth. Three test cases is a small sample, and the skill here is a prompt and a set of reference notes, so the numbers describe the behaviour of a model following instructions rather than the quality of any investment decision.

A prompt pack, with an unclear licence and a quiet branch

Two practical points sit in the metadata rather than the README text. The repository does not declare a licence, and it has no GitHub releases, so there is no versioned artefact to pin and no stated grant to rely on if you intend to redistribute the skill inside a commercial product. The last push was on 2026-04-13, and the repository is not archived, so the work is recent but the branch is not moving quickly. The scope of the repository is also worth noticing, because the topics include bitcoin, ethereum and finance alongside buffett and ai, which suggests the collection is intended to grow beyond equities. For a reader, the practical value of this project is not that the analysis will be right, since nothing in it reads a filing or a price, but that the questions you ask get asked the same way every time, with the same sections and the same explicit statement of what the decision depends on.

Editorial conclusion

Adopt buffett-skills when you want every stock or company question answered through the same checklist and a fixed report structure, since the skill ships eight reference files, a dispatch table and a template that requires every section to be filled. Do not adopt it as a source of investment advice or a data source, because it is a prompt and reference pack with no market data, no filings and no prices of its own, and the evaluation it reports covers three test cases. Verify first that your agent reads the reference files it needs, and read the eval table in the README yourself, because the pass rate gain is reported alongside a large increase in tokens and time.

Frequently asked questions

Official sources

1. agi-now/buffett-skills on GitHub
2. Issues
3. Project website
4. README