How Splash works

An agent system for charts, maps and video

Most researchers and investigators know a finding would land better as a chart or a map. What they lack is the time and the craft to build one. Splash is seventeen skills that teach an AI assistant that craft — how to read your evidence, propose a form that the data actually supports, draw it for real, and hand you something a CMS will take.

It is local-first by design. Stories and your own configuration live outside the replaceable Splash checkout, credentials stay in the operating system's protected store through Engine, and every delivery is a file or a directory you control. Nothing is uploaded to run it.

Six moves and two doors that only you open

The shape below is the whole product. What matters is not the order of the steps — it is the two gates. Splash never advances past a storyboard you have not approved, or delivers a render you have not seen.

Why the gates are hard stops

Craft-skill selection is ranked, but a ranking is a suggestion. The journalist's confirmed takeaway always wins, and Splash refuses to treat an unapproved storyboard as settled. If your data will not support the form you asked for, it says so at the storyboard rather than drawing something the evidence does not carry.

Sixteen skills, composed by one

You invoke splash. It recovers a story's phase from the story's own directory and calls the rest in order. The others are here because you can also call them directly when you know exactly what you want.

The same story in four formats

A beat is built once. The format decides how it leaves, and each route carries only the delivery forms it can actually honour.

Forty-four forms sit behind those routes — charts, maps and scrollies. Splash knows what each one needs, and says so when your data will not support the one you asked for.

Everything lands in a folder you own

Splash does not hold your work. Source files, editorial decisions, rendered drafts, approvals and delivery artifacts all sit in a journalist-owned story directory, outside the Splash checkout — so you can replace or delete Splash without losing a story.

stories/<your-story>/
  source/          the frozen article and data, never edited in place
  STORYBOARD.md    the angle, the forms, and what you approved
  beats/           one directory per beat, with its rendering code
  renders/         the real output, versioned per approval round
  delivery/        what you hand to the CMS

Two ways to get it

Managed — recommended

Indicator Labs on Mac or Windows, with guided setup, credential management, repair and automatic updates. Join at buriedsignals.com/join →

Agent-led manual install

The same Engine, through its public signed standalone release. Fetch the bootstrap JSON for your platform, verify both the SHA-256 and the Minisign signature before extracting or executing bsig, then sync the catalog.

https://navigator.indicator.media/api/artifacts/bootstrap/bsig/<platform>
# darwin-arm64 · darwin-amd64 · linux-arm64 · linux-amd64 · windows-amd64

bsig catalog sync
bsig plan install splash

Windows 11 amd64 is a prepared host: Engine does not install Goose Desktop, Git, Bun, Docker or Node, and anything missing fails at plan time. Review the emitted plan, apply that exact plan.json, and launch Goose from the Start menu once doctor is green. Do not run install.sh on Windows.

Common questions

Deeper docs on GitHub

The repository carries the doctrine reference, the visual catalogue that defines all forty-four forms, the skills themselves, and the worked stories used to test them.