Skip to main content
The skill is a set of instructions you give Claude alongside the Semilattice MCP connection. It tells Claude how to turn a document into simulations: which claims in it are worth testing, which user model should answer each one, what the whole job would cost before anything runs, and how to report what came back.

Semilattice plugin

The skill and the MCP connection, in one install for Claude.ai, Cowork and Claude Code.

SKILL.md

The instructions themselves, to read or paste into another client.
The skill calls Semilattice MCP tools. The plugin connects the MCP server for you; anywhere else, connect it first. See Connect to Semilattice MCP.

​
What the skill does

​
Matching the strategy to the request

The skill picks the narrowest strategy your words support: When more options could be written, Claude offers the next strategy up with its cost in simulations and time, and you decide.

​
Decomposition

The skill reads your input as a list of claims about users, including the ones a document assumes rather than states: that users want this kind of product at all, or that the competitor it names is the one users would otherwise pick. Where a document has already settled on an option, the claim it tests is that the chosen option beats the ones it passed over.

​
Pricing before anything runs

Before creating a simulation, Claude tells you how many testable claims your input contains, how many simulations that is, and roughly how long they take. You choose how much of it to run, and one yes covers the batch you agreed to. The claims you leave out are listed at the end so you can pick them up later.

​
What a survey question can be

Every survey question is a single-choice Likert-scale item, and the whole modelled audience answers all of them. A simulation holds up to 40 questions. There is no conditional logic, no subgroup filtering, no cross-tabulation. The skill plans inside those limits, and where a claim has to be reworded to fit them it says so and gives you the reworded version.

​
Choosing who answers

Each claim is tested against one user model, matched to the audience the claim is about. Where your organisation has its own user models, built from your own data, those answer claims about your users, and the public models cover audiences yours do not. A single claim is not run against several models at once. A claim about part of an audience, such as one age band or region, runs on a segment of the model that covers it.

​
Install

The skill ships in the Semilattice plugin, which also connects the MCP server.
  1. Add the Semilattice plugin from the plugin directory
  2. Open the plugin’s Connectors tab and connect Semilattice
  3. Sign in with your Google email
🎉 Start a new conversation and the skill will load whenever you ask Claude to test something with users.