# PortCo Desk

> One portfolio company's period reporting pack, worked four ways in a single sitting:
> where it stands against the plan it was underwritten on, whether the unit economics
> support that plan, what closes the gap, and what the exit looks like on this trajectory.

URL: https://portco-desk.skillsafe.ai/
API tutorial: https://portco-desk.skillsafe.ai/api.html
Token panel: https://portco-desk.skillsafe.ai/tokens.html
Platform: SkillSafe apps (https://skillsafe.ai/)
Model: gpt-terra (currently resolves to gpt-5.6-terra) · publisher markup 1000 bps
Cost: signing in is required to run a lane; the whole prescan and every prescan export is free.

## What it is for

A deal team, an operating partner or a finance lead receives one portfolio company's monthly
or quarterly reporting pack. PortCo Desk is the sitting that follows: read the pack against
plan, understand the unit economics underneath it, decide what to do about the gap, and
re-price the exit.

It is analysis of pasted figures. It is **not** investment advice, a valuation, a fairness
opinion or an audit, and it never recommends buying, selling or holding anything.

## The four lanes

The app has one system prompt and one model, and routes on an explicit `task` field.

| `task` | Name | The question it answers | Its extra output block |
| --- | --- | --- | --- |
| `monitor` | Monitor | Where is the company against the plan it was underwritten on, and what is driving each gap? | `drivers` |
| `unit` | Unit economics | Do the unit economics support the plan, and which ratio is load-bearing? | `metrics` |
| `plan` | Value creation | What closes the gap, and what happens in the first hundred days? | `levers` |
| `returns` | Exit case | On this trajectory, what does the exit look like, and does it clear the hurdle? | `cases` |

The intended pipeline is `monitor` → `unit` → `plan` → `returns`, and each result carries a
digest of its conclusions into the next lane as `carryover`. Every lane stands alone: opening
`returns` first still returns a complete answer.

## The input contract

One JSON object. `task` is the only field that changes the contract.

- `task` — `monitor` | `unit` | `plan` | `returns`
- `pack` — the reporting pack as text: a metrics table of actuals against plan, usually with a
  prior-year column. The primary input for every lane.
- `deal` — the deal terms as `key: value` lines. Optional, but the exit case needs them.
- `stage` — `buyout` | `growth` | `carve-out` | `platform` | `venture` | `unknown`
- `emphasis` — `general` | `revenue` | `margin` | `cash` | `exit`
- `context` — free text about the period. Attributed by the model, never treated as a figure.
- `carryover` — a digest of the previous lane's conclusions. Empty on a first run.
- `prescan_facts` — the browser's own arithmetic, handed over as ground truth: `stats`,
  `variances`, `derived`, `returns`, `sensitivity` and `flags`.

The object is the request body itself. There is no `input` wrapper — and note that sending one
anyway is **not** rejected: `{"input": {...}}` returns 200 with a plausible `hold_credits` (1,719 against the
correct body's 1,833),
because the platform passes the object straight through, so the model never sees `task` and
answers whichever lane it guesses. Verified live on 2026-08-19. The same is true of an invented
app-slug header: a bogus value returns 200 and changes nothing. There is no error to catch.

## The output contract

One JSON object, identical in every lane apart from the extra block:

`lane`, `title`, `company`, `period`, `verdict` (`on-track` | `watch` | `at-risk`), `headline`,
`summary`, `checks[]` (exactly the lane's eight named checks, in order, each `pass` | `warn` |
`fail` | `unknown`), `findings[]` (`id` PD-nnn, `severity`, `area`, `metric`, `detail`,
`so_what`, `action`), `coverage_check[]` (one entry per prescan flag id, `confirmed` or
`set-aside`), `data_gaps[]`, `next_steps[]`, `disclaimer`, plus the lane's own block.

## What the browser does for free, before any model call

This is most of the app, and it costs nothing:

- **Reads the table.** The delimiter (pipe, tab, semicolon, comma, aligned space) is chosen on
  evidence across the whole paste, not off the first line. Quoted cells survive. Rule rows are
  ignored. The actual, plan and prior-year columns are resolved from the header; where there is
  no header the assumption is made explicit rather than hidden.
- **Reads the figures the way packs write them.** `(1,234)` is negative, `$1.2m` is 1,200,000,
  `12.4%` is a percentage, `8.5x` is a multiple, `1 234,56` is a European decimal.
- **Knows over fifty standard line items** whatever the pack calls them — `S&M`, `S&M spend`,
  `Total S&M costs` all resolve to one id — and knows which direction each should move. A churn
  line under plan is favourable; a revenue line under plan is not.
- **Varies percentage lines in percentage points**, not as a percentage of a percentage.
- **Derives every unit economic the rows allow**: gross margin, EBITDA margin, ARPA, CAC, LTV,
  LTV/CAC, CAC payback, NRR, GRR, net new ARR, growth, Rule of 40, burn multiple, magic number,
  cash conversion, net leverage. Each records the rows it rests on and the band it is judged
  against. A ratio whose inputs are absent is reported as **not derivable** and never estimated.
- **Completes the entry triangle** — EV, EBITDA, multiple, equity, net debt — wherever two sides
  are known, and states each derivation.
- **Solves the IRR by bisection** over the annual flow vector, reporting the iteration count and
  the residual, and says what one more year of hold costs at the same money multiple.
- **Attributes the value-creation bridge** to EBITDA growth, multiple change and debt paydown,
  and reports the unexplained residual rather than absorbing it into one of the three.
- **Computes a five-by-five exit sensitivity grid** over exit multiple and exit EBITDA, every
  cell computed rather than interpolated.
- **Raises deterministic flags** the model must then reconcile one by one: a margin that does not
  recompute from the lines above it, an ARR movement that does not close to closing ARR, a plan
  column that matches the actuals on most rows, receivables outgrowing revenue, an EBITDA beat
  that did not reach free cash flow, a covenant with no headroom, a CAC payback longer than the
  hold that remains, negative net new ARR, and multiple expansion presented as a result rather
  than an assumption.

## How the model is held to it

Every prescan flag must come back either `confirmed` or explicitly `set-aside`, and the page
names any flag the model dropped and any id it invented. The unit lane must judge every ratio
the browser derived, and the page reports any it skipped. The returns lane's base case must
reproduce the browser's own MOIC and IRR, and the page checks that it did. Checks are rendered
against the lane's exact roster, so a check the model skipped shows as *not returned* rather
than shortening the table.

## Sources

A derived work of four agent skills published in the Anthropic financial-services plugin
repository, credited in the app footer and in its description:

- `@anthropics/portfolio-monitoring` — the `monitor` lane
- `@anthropics/unit-economics` — the `unit` lane
- `@anthropics/value-creation-plan` — the `plan` lane
- `@anthropics/returns-analysis` — the `returns` lane

Not affiliated with those skills' authors, and not endorsed by them.

## Privacy

The pack is read in your browser. Nothing is uploaded until you run a lane, and then only for
that one analysis. History is stored in the declared `analyses` collection on your own SkillSafe
account, mirrored in this browser for instant paint.