How to use the Investment Thesis Drift Checker agent skill

What the thesis-drift skill makes the agent do

An agent skill is a folder with a SKILL.md file that tells an AI agent a specific workflow to follow when you ask for that kind of task. thesis-drift teaches the agent to check whether an investment thesis has drifted, meaning whether the evidence behind a long-term holding has changed in a way that matters.

The skill is built around one rule from its source: only admit a thesis change when the evidence changes. Rewording a report is not drift, and a falling share price does not by itself mean the business got worse.

• Separates three things: factual change (revenue, margins, cash flow, governance, competition, capital allocation), price change (sentiment or valuation multiples moving while the business stays the same), and wording change (different phrasing with the same evidence).
• Compares the two reports on fixed dimensions: valuation anchor, core assumptions, red-line list, management quality and competitive moat.
• Labels each dimension as Improved, Unchanged or Weakened, and requires a specific piece of new evidence for every change it reports.
• Asks for exact arithmetic through tools/financial_rigor.py rather than mental math, and flags numbers that cannot be cross-checked as low confidence.

Install the thesis-drift skill

Download the ZIP for this skill from this page and unzip it. The skill is a single file, SKILL.md, so the folder you unzip should contain that file. Then choose one of the locations below.

For Claude Code, put the skill folder in your user skills directory at ~/.claude/skills/, or in a project's .claude/skills/ folder if you want it available only inside one repository. For the Claude apps, upload the ZIP under Settings → Capabilities → Skills.

• Personal (all projects): unzip into ~/.claude/skills/
• Single project: unzip into the project's .claude/skills/ folder
• Claude apps: upload the ZIP under Settings → Capabilities → Skills
• Codex users: the file is a Codex adapter of the same workflow and is intended to run from a repository that contains tools/ and AGENTS.md

Check the requirements before you run it

This skill is not a standalone analyst. It depends on a specific repository layout and on prior reports. Check these before your first run so you do not hit a dead end halfway through.

• Python 3, because the workflow calls tools/financial_rigor.py for calculations. That script is not included in the single file on this page, so it must come from the AI Berkshire repository.
• The AI Berkshire repository structure, with a tools/ folder and an AGENTS.md file present.
• Prior thesis reports saved in a reports/ folder. These are ideally produced by a thesis-tracker workflow, which the skill references for its structured sections.
• Shell access, so the agent can run the date command and Python scripts.
• Language: the instructions are written in Chinese and refer to Chinese-market filings, so the reports you feed in should be in a form the agent can read.

Trigger the skill with natural requests

You do not need special syntax beyond naming the skill and giving it the company and the reports. The skill accepts three input patterns: two explicit report paths, a company name alone (which makes it look for reports/{company}-thesis.md and dated snapshots beside it), or a company name with dated report paths. The source's own examples use the /thesis-drift command form with these arguments.

If you only have one report, or no earlier snapshot, the skill stops and does not invent a baseline. It tells you to build one first with the thesis-tracker workflow.

• Compare two reports: /thesis-drift 公司名 reports/公司名-thesis-旧日期.md reports/公司名-thesis-新日期.md
• Let the skill find snapshots: /thesis-drift 公司名
• Ask after new earnings: run the drift check on the report saved before the last results and the one saved after them
• If no baseline exists: the skill will recommend running /thesis-tracker 公司名 first

Read the output report

The skill writes a drift report with fixed sections: the comparison objects and time span, an overall verdict on whether the thesis drifted, a dimension table, a list of evidence differences, valuation and numeric checks, how the suggested action should move, open uncertainties, and what to watch next time.

The overall verdict must answer five questions: has the thesis drifted, where the drift came from, whether it was a factual change or a price change, how the suggested action (for example Watch to Buy, or Hold to Reduce) should change, and what evidence is still needed.

• Unchanged rows leave the trigger evidence blank rather than inventing support.
• Every Improved or Weakened call must point to a specific source such as a financial line item, a regulatory filing, a news event or a valuation calculation.
• Red-line triggers take priority over a cheap valuation, so a low P/E cannot hide a triggered red line.