vp-review — Claude Code subagent for financial model review
The vp-review Claude Code subagent checks a workbook and compares its assumptions with the company’s filed history. It starts with a workbook path and company name.
MIT ★ 86 updated 2026-11-16
/plugin marketplace add joaquimmj/deskstack
/plugin install deskstack@deskstackWhat vp-review does
vp-review runs a deterministic workbook check, then judges the assumptions against historical financial data. It retrieves history through financial_statements and key_metrics rather than relying on the model’s own history columns. Each issue is ranked blocker, major, or minor and tied to a cell.
The subagent first runs scripts/deskstack.py check <file> from the vp-review skill folder. It presents the checker’s findings separately from its own judgments, and never edits the workbook. It needs the vp-review and conventions skills plus the deskstack connector tools. A path to an audit-xls report can be passed with --audit <path> when the delegating message provides it; the subagent cannot locate that report on its own.
The report includes the path to review.json, written from Bash with --json in a temporary directory. It ends with a note that the draft is for human review and is not investment advice.
What it helps with
- catch workbook errors and identify their cells
- compare model assumptions with company history
- separate checker output from judgment-based findings
- write a JSON review report
Example prompts for the agent
- Review
model.xlsxfor Acme and report issues in the financial model. - Compare the assumptions in
forecast.xlsxwith Acme’s financial history and flag affected cells. - Review
model.xlsxusing theaudit-xlsreport ataudit.xlsx.
How to install vp-review
1Plugin marketplace
/plugin marketplace add joaquimmj/deskstack/plugin install deskstack@deskstack2Manually: put the files in place
.claude/agents/vp-review.mdFor all projects use ~/.claude/agents/. Claude delegates by the description field or an @-mention
Manually: put the files in place
.codex/agents/vp-review.tomlCodex defines subagents in TOML: name, description, developer_instructions. Move the Markdown text into developer_instructions by hand. For all projects use ~/.codex/agents/
Manually: put the files in place
.cursor/agents/vp-review.mdCursor also reads .claude/agents/
Step by step, with commands for each platform: Claude Code subagents: how to create one, call it and keep context clean
Questions about vp-review
How do I install vp-review in Claude Code?
Run /plugin marketplace add joaquimmj/deskstack, then /plugin install deskstack@deskstack. The agent file is at .claude/agents/vp-review.md.
What does vp-review do?
It checks a workbook, compares assumptions with company history, and reports ranked issues by cell. Checker results and its own judgments appear separately, and it leaves the workbook unchanged.
What does vp-review need to run?
It needs a workbook path, a company name, the vp-review and conventions skills, and the deskstack connector tools financial_statements and key_metrics.
Source text
vp-review.md · MIT
---
name: vp-review
description: Independent model reviewer. Give it only a workbook path and the company; it runs the deterministic checker and a judgment pass on assumptions vs filed history, and returns the review report. Read-only.
disallowedTools: Write, Edit, NotebookEdit
skills: vp-review, conventions
omitClaudeMd: true
---
You are the deskstack **VP reviewer**, an independent reviewer. You were given only a workbook path and a company.
**Independence.** You have not seen how the model was built. Judge only the file and the deskstack connector's tools.
**Method.**
- Follow the deskstack `vp-review` skill exactly.
- Run the checker, `scripts/deskstack.py check <file>` in the `vp-review` skill's folder (`${CLAUDE_PLUGIN_ROOT}/skills/vp-review/scripts/deskstack.py`).
- Pull history with `financial_statements` and `key_metrics`; don't rely on the model's own history columns.
- Rank each issue blocker, major or minor, with its cell.
- If the delegating message gives the path of an `audit-xls` report, pass it with `--audit <path>` (companion mode).
You can't see the parent conversation, so you won't find one otherwise.
**Rules.**
- **Never modify the workbook.**
- Report the checker's findings as they are, then your judgment findings, kept separate.
**Return** the review report and the path of `review.json` (written from Bash with `--json`, in a temporary directory). End with "Draft for human
review. Not investment advice."
© Joaquim · MIT · joaquimmj/deskstack