# Three-Statement Desk

> A linked three-statement financial model that runs in the browser from pasted historical
> statements, followed by two paid steps: propose the drivers from the history and your guidance
> notes, then review the finished model. Every number either step writes is checked against the
> model. https://three-statement-desk.skillsafe.ai/

Three-Statement Desk is a web app on SkillSafe derived from the agent skill
@anthropics/3-statement-model (anthropics/financial-services-plugins, Apache-2.0). It runs on
gpt-terra and is metered per step; the model itself is free and needs no account.

## What the free model computes

- Reads pasted statements (CSV, TSV, Markdown pipe tables or aligned columns; units, thousands or
  billions, converted to $ millions), maps each row to a line item and lists the rows it could not
  place.
- Rebuilds each historical year, derives the few lines that can be derived exactly, and checks that
  each balance sheet balances and that net income ties to its components.
- Historical ratios per year and averaged: growth, gross and EBITDA margin, operating expenses,
  D&A and capex as a share of revenue, receivable, inventory and payable days, other working
  capital, tax rate, interest rate, payout.
- A linked projection of 1 to 10 years: interest on beginning balances (no circular reference),
  PP&E roll-forward, working capital from days, cash from operations, capex, scheduled debt
  repayment, dividends and buybacks, cash as the balancing item and a revolver drawn below the
  minimum cash, capped at its commitment.
- Integrity checks every year: the balance sheet balances, cash ties to the cash flow statement,
  equity and PP&E roll forward, no negative balances.
- Flags: liquidity shortfall, negative equity, negative EBITDA, losses, revolver draws, growth and
  margins above history, working-capital day shifts, leverage above 4.0x, interest cover under
  2.0x, capex below D&A, payouts above earnings, tax-rate shifts, idle cash.
- Baseline compare: pin one case, change a driver, and see what moved (final-year revenue and
  net income, CAGR, EBITDA margin, cumulative cash flow, lowest cash, peak revolver, debt,
  leverage, interest cover, failing checks, flags) with the list of changed inputs.
- Exports: model CSV, Markdown, tab-separated copy that pastes straight into spreadsheet cells.

## The two paid steps (field `task`)

- `drivers`: one proposal per driver - value, basis (guidance, history or judgement), rationale and
  a verbatim quote from the notes for guidance - plus cautions, questions and a summary. The page
  checks every quote against the notes and applies the proposals in one click.
- `review`: verdict (sound, needs_work, not_credible), headline, what the model depends on,
  findings (severity, area, driver, issue, fix), one response per flag, a checks note, questions and
  a summary. The verdict is checked against the integrity checks and flag severities.

## Limits

It does not model acquisitions, new debt issuance, leases, stock compensation, deferred taxes, loss
carryforwards or quarterly timing; other long-term assets and liabilities are held flat. It does
not know market data or real companies. It is analysis, not investment advice.

## Links

- App: https://three-statement-desk.skillsafe.ai/
- API tutorial: https://three-statement-desk.skillsafe.ai/api.html
- Source skill: https://skillsafe.ai/skill/@anthropics/3-statement-model