How to install AI Berkshire: Use Codex/Claude Code to do multi-agent value investment research

AI Berkshire installation and usage tutorial: Deploy value investing skills in Codex or Claude Code, run company research, financial report analysis and multi-agent processes, and control data and investment risks.

This tutorial by AI Berkshire focuses on the specific task in the title. Multi-agent investment research amplifies data errors and shared biases; any valuation, holding, or trading recommendations must go back to earnings reports, market data timestamps, and clear assumptions.

Below is a demonstration using a test environment and revocable data. When accounts, keys, machine control, investments or emergency communications are involved, manual confirmation must be done.

First position the project as a research demonstration

This section only deals with “positioning the project as a research demonstration first”. Save the current state first and then run the minimal test. Do not change multiple conditions at the same time.

1
git status --short

For AI Berkshire, results must be judged within this boundary: multi-agent investment research amplifies data errors and common biases; any valuation, positioning or trading recommendations must go back to financial reports, market data timestamps and clear assumptions.

If the results are not repeatable, record the version, input, and logs, then narrow the scope of the test.

Check source and license before installing Skill

It is recommended to execute in the following order:

  1. Check the actual version and source.
  2. Use a test account, test directory, or test device.
  3. Perform an undoable action.
  4. Save the output, exit code, or screenshot.
  5. Revert the changes and re-verify.
1
Get-Command codex, claude -ErrorAction SilentlyContinue

The completion standard is to get verifiable results by “checking the source and license before installing the skill”, rather than the command just not reporting an error.

Only enable one research role for the first time

Check items Passing criteria Stop signal
Only one research role is enabled for the first time Input, output and permission scope are clear The tool automatically expands the scope of operations
Source Can return to the official warehouse or original data Rely on unknown mirrors or second-hand conclusions
Security No secret value in log Token, cookie or private data present
Recovery Ability to return to pre-execution state No backup or undo methods
1
# 先运行财务数据核对角色

When a stop signal appears, resume first and do not continue to the next step.

Ticker and market code should be clear

This section only deals with “Ticker and market codes must be clear”. Save the current state first and then run the minimal test. Do not change multiple conditions at the same time.

1
# 记录交易所、币种和证券类型

For AI Berkshire, results must be judged within this boundary: multi-agent investment research amplifies data errors and common biases; any valuation, positioning or trading recommendations must go back to financial reports, market data timestamps and clear assumptions.

If the results are not repeatable, record the version, input, and logs, then narrow the scope of the test.

Financial report data must include the reporting period

It is recommended to execute in the following order:

  1. Check the actual version and source.
  2. Use a test account, test directory, or test device.
  3. Perform an undoable action.
  4. Save the output, exit code, or screenshot.
  5. Revert the changes and re-verify.
1
# 保存 fiscal period 与 filing date

The completion standard is that “financial report data must be with the reporting period” to obtain verifiable results, rather than the order just not reporting errors.

Market prices must have timestamps

Check items Passing criteria Stop signal
Market prices must be timestamped Input, output and authority scopes are clear Tools automatically expand the scope of operations
Source Can return to the official warehouse or original data Rely on unknown mirrors or second-hand conclusions
Security No secret value in log Token, cookie or private data present
Recovery Ability to return to pre-execution state No backup or undo methods
1
# 保存时区、收盘价类型与数据源

When a stop signal appears, resume first and do not continue to the next step.

Value investment assumptions are written as modifiable parameters

This section only deals with “value investing assumptions written as modifiable parameters”. Save the current state first and then run the minimal test. Do not change multiple conditions at the same time.

1
# 增长率、折现率、安全边际

For AI Berkshire, results must be judged within this boundary: multi-agent investment research amplifies data errors and common biases; any valuation, positioning or trading recommendations must go back to financial reports, market data timestamps and clear assumptions.

If the results are not repeatable, record the version, input, and logs, then narrow the scope of the test.

Multiple Agents should not copy each other’s conclusions

It is recommended to execute in the following order:

  1. Check the actual version and source.
  2. Use a test account, test directory, or test device.
  3. Perform an undoable action.
  4. Save the output, exit code, or screenshot.
  5. Revert the changes and re-verify.
1
# 要求每个角色提供独立证据

The completion standard is “multiple agents should not copy each other’s conclusions” to obtain verifiable results, rather than the command just not reporting an error.

The final aggregator should preserve differences

Check items Passing criteria Stop signal
The final aggregator needs to preserve differences Input, output, and permission scopes are clear Tools automatically expand the scope of operations
Source Can return to the official warehouse or original data Rely on unknown mirrors or second-hand conclusions
Security No secret value in log Token, cookie or private data present
Recovery Ability to return to pre-execution state No backup or undo methods
1
# 不用多数票隐藏少数风险

When a stop signal appears, resume first and do not continue to the next step.

Token consumption is controlled from role and context

This section only deals with “Token consumption from role and context control”. Save the current state first and then run the minimal test. Do not change multiple conditions at the same time.

1
# 先限制角色数、历史长度和工具调用

For AI Berkshire, results must be judged within this boundary: multi-agent investment research amplifies data errors and common biases; any valuation, positioning or trading recommendations must go back to financial reports, market data timestamps and clear assumptions.

If the results are not repeatable, record the version, input, and logs, then narrow the scope of the test.

It is prohibited to give brokerage account keys to the research process

It is recommended to execute in the following order:

  1. Check the actual version and source.
  2. Use a test account, test directory, or test device.
  3. Perform an undoable action.
  4. Save the output, exit code, or screenshot.
  5. Revert the changes and re-verify.
1
git grep -n -I -E 'API_KEY|TOKEN' -- .

The completion standard is “forbidden to hand over brokerage account keys to the research process” to obtain verifiable results, rather than the command just not reporting an error.

Why are the results different when running the same problem repeatedly?

Check items Passing criteria Stop signal
Why are the results different when running the same problem repeatedly Clear input, output and authority scope The tool automatically expands the scope of operations
Source Can return to the official warehouse or original data Rely on unknown mirrors or second-hand conclusions
Security No secret value in log Token, cookie or private data present
Recovery Ability to return to pre-execution state No backup or undo methods
1
# 固定模型、温度、数据快照和提示词

When a stop signal appears, resume first and do not continue to the next step.

Review an estimate using manual calculations

This section only deals with “reviewing a valuation using manual calculations”. Save the current state first and then run the minimal test. Do not change multiple conditions at the same time.

1
# 独立计算自由现金流与折现

For AI Berkshire, results must be judged within this boundary: multi-agent investment research amplifies data errors and common biases; any valuation, positioning or trading recommendations must go back to financial reports, market data timestamps and clear assumptions.

If the results are not repeatable, record the version, input, and logs, then narrow the scope of the test.

Check if the quote actually supports the conclusion

It is recommended to execute in the following order:

  1. Check the actual version and source.
  2. Use a test account, test directory, or test device.
  3. Perform an undoable action.
  4. Save the output, exit code, or screenshot.
  5. Revert the changes and re-verify.
1
# 打开原始 filing 对照数字

The completion standard is to “check whether the reference really supports the conclusion” and obtain verifiable results, rather than the command just not reporting an error.

Save research reports instead of automatically placing orders

Check items Passing criteria Stop signal
Save research reports instead of automatically placing orders Clear input, output and authority scope Tools automatically expand the scope of operations
Source Can return to the official warehouse or original data Rely on unknown mirrors or second-hand conclusions
Security No secret value in log Token, cookie or private data present
Recovery Ability to return to pre-execution state No backup or undo methods
1
git status --short

When a stop signal appears, resume first and do not continue to the next step.

Final disclaimer is not a substitute for verification

This section only deals with “Final disclaimer is not a substitute for verification”. Save the current state first and then run the minimal test. Do not change multiple conditions at the same time.

1
# 明确教育研究用途与数据限制

For AI Berkshire, results must be judged within this boundary: multi-agent investment research amplifies data errors and common biases; any valuation, positioning or trading recommendations must go back to financial reports, market data timestamps and clear assumptions.

If the results are not repeatable, record the version, input, and logs, then narrow the scope of the test.

Role prompt words are also experimental configurations

It is recommended to execute in the following order:

  1. Check the actual version and source.
  2. Use a test account, test directory, or test device.
  3. Perform an undoable action.
  4. Save the output, exit code, or screenshot.
  5. Revert the changes and re-verify.
1
git diff -- .

The completion standard is that “the character prompt word also belongs to the experimental configuration” and a verifiable result is obtained, rather than that the command happens to not report an error.

How to unify cross-currency reports of the same company

Check items Passing criteria Stop signal
How to unify cross-currency reports of the same company Clear input, output and authority scope The tool automatically expands the scope of operations
Source Can return to the official warehouse or original data Rely on unknown mirrors or second-hand conclusions
Security No secret value in log Token, cookie or private data present
Recovery Ability to return to pre-execution state No backup or undo methods
1
# 记录汇率日期与换算方法

When a stop signal appears, resume first and do not continue to the next step.

Preserve the opposing Agent’s objections to the report

This section only deals with “preserving the opposing Agent’s objections until reporting”. Save the current state first and then run the minimal test. Do not change multiple conditions at the same time.

1
# 单列未解决风险与证据缺口

For AI Berkshire, results must be judged within this boundary: multi-agent investment research amplifies data errors and common biases; any valuation, positioning or trading recommendations must go back to financial reports, market data timestamps and clear assumptions.

If the results are not repeatable, record the version, input, and logs, then narrow the scope of the test.

Final Reconciliation of AI Berkshire

  • The versions and sources used are documented.
  • Test input and formal data have been isolated.
  • Commands, configurations and actual output can correspond.
  • The failed path was actively tested at least once.
  • Keys, accounts and private data do not enter Git or logs.
  • There are clear actions for upgrading, stopping and resuming.

If you still can’t explain how the results came about, keep AI Berkshire in a test environment without extending permissions or replacing existing stable processes.