# Carto Desk

> Paste a Mapbox GL style specification document and review it four ways in one sitting: an audit
> against the style spec, a cartographic critique, a data-visualization review of the data-driven
> layers, and a render-cost pass. A free in-browser prescan parses the style first and the model is
> held to what it found.

Site: https://carto-desk.skillsafe.ai/
API tutorial: https://carto-desk.skillsafe.ai/api.html
Token panel: https://carto-desk.skillsafe.ai/tokens.html (noindex)

## What it is for

One work object: a Mapbox GL style JSON, style specification version 8. One user, one sitting, four
lanes over that same document — inspect it, decide how it should look, fix the data encoding, then
price what it costs to draw. Each lane's result carries into the next through a handoff button.

## The lanes

| `task` | Lane | Answers | Source skill |
| --- | --- | --- | --- |
| `audit` | Audit | Does this style parse, resolve and ship? | `@mapbox/mapbox-style-quality` |
| `carto` | Cartography | Will a reader find what they came for? | `@mapbox/mapbox-cartography` |
| `dataviz` | Data layers | Are the data layers encoding the data honestly? | `@mapbox/mapbox-data-visualization-patterns` |
| `perf` | Render cost | What does this style cost to draw? | `@mapbox/mapbox-web-performance-patterns` |

## What runs for free in the browser, before any model call

A real style reader, not a regex pass. It parses the JSON and reports:

- layer and source integrity — every layer's `source` resolved against `sources`, `source-layer`
  required on vector sources, duplicate layer ids, unknown layer types, unused sources
- paint versus layout placement — a layout property in `paint` (or the reverse) is silently ignored
  by the renderer, so the style appears to work while the property does nothing
- sprite and glyphs — `icon-image` with no `sprite`, `text-field` with no `glyphs`
- zoom ranges — `minzoom` >= `maxzoom` means the layer never draws
- embedded credentials — a Mapbox secret token (`sk.`) in a served style is critical; a public token
  (`pk.`) belongs in the client's init call with URL restrictions
- WCAG label contrast — relative luminance and contrast ratio computed from the declared text, halo
  and background colours
- layer hygiene and cost — hidden layers, identical layer bodies, legacy `{property, stops}`
  functions, ungated expensive layers, symbol-layer count against a draw budget

Those findings are sent to the model as `prescan_facts`, and the model must reconcile every one of
them — confirmed or set aside — in `coverage_check`. The app shows any it skipped.

## Input contract

```json
{
  "task": "audit | carto | dataviz | perf",
  "style": "the Mapbox GL style JSON, as text",
  "usage": "what the map is for; GL JS init code welcome for the perf lane",
  "audience": "general | analyst | field | print | unknown",
  "emphasis": "general | legibility | accessibility | data | speed",
  "carryover": "a digest of a previous lane's review, or an empty string",
  "prescan_facts": {"stats": {}, "flags": [{"id": "SC-001", "title": "", "severity": "", "layer": "", "detail": ""}]},
  "retry_note": "present ONLY on the app's reformat retry: a correction directive about the model's own previous malformed reply, never user content"
}
```

`retry_note` is optional and absent from a first attempt. When the app cannot parse a reply it
retries once with `retry_note` set to what was wrong with that reply. It is an instruction from the
app to the model about its own previous output — it is never something the user wrote, and nothing in
it describes the style.

## Output contract

One JSON object. The envelope is identical in every lane:
`lane`, `title`, `style_name`, `verdict` (`ship` | `fix-first` | `rework`), `headline`, `summary`,
`checks[]` ({name, status, note} — the lane's named checks, in order), `findings[]`
({id `CD-nnn`, title, severity, area, layer, detail, fix, snippet}), `coverage_check[]`
({prescan_id, status, note}) and `next_steps[]`.

Each lane adds exactly one block: `audit` → `blocking[]`; `carto` → `palette[]`;
`dataviz` → `recipes[]`; `perf` → `budget{}` plus `wins[]`.

## Model and cost

Model `gpt-terra` (currently `gpt-5.6-terra`), publisher markup 1000 bps. `estimate` is free and is
re-taken on every lane switch, because the hold differs per lane. Four of the eight example×lane
combinations replay a saved run at zero cost without an account: the delivery ops style for `audit`,
`carto` and `perf`, and the air-quality style for `dataviz`. The other four are ordinary paid runs,
and the lane bar marks which is which for the style currently loaded. Guest runs are not sponsored,
so every non-replay run needs a signed-in account.

## Sources and attribution

Carto Desk is a derived work built on four skills from `@mapbox/mapbox-agent-skills`:
`@mapbox/mapbox-style-quality`, `@mapbox/mapbox-cartography`,
`@mapbox/mapbox-data-visualization-patterns`, `@mapbox/mapbox-web-performance-patterns`.
Mapbox and Mapbox GL JS are trademarks of Mapbox, Inc. This app is not affiliated with or endorsed
by Mapbox. Nothing about a pasted style is stored anywhere except the user's own SkillSafe account
history.