Answer your evaluation questions from your documents, inside your own Claude

What you get

One report per question, answer first. Where questions share a step, such as one coding of the documents, they count from the same work, so the answers keep the same definitions. Click any number, source ID or quotation to see the passages behind it. It ends with how it was made: the methods annex a commissioner asks for. A Word copy comes with it.

Example reports: household health, five questions in one combined report, AI at work, MECC training.

How a run goes

1. Attach your documents and type /rubicon with your question.
2. Confirm the documents Claude found.
3. Agree an overall design: what other questions you have for these documents, what analysis they can share, and the order to answer them in. You can plan them all now, let the outline grow, or take one question at a time.
4. Agree each question before anything is read: what the answer is for, which groups to compare.
5. Claude carries out the workflow the question needs, step by step: for instance coding the passages on each part of it, counting and comparing groups with code, and judging the counts against a rubric.
6. A second coder codes again without seeing the first. A fresh reader settles disagreements and checks every sentence and quotation.
7. Code creates the report.

A run somebody else can rebuild

Beside the report is a zip of the whole run: your documents, the coded passages and the workflow, in an open format. Anyone can recount every number from it. The report, its Word copy and the zip end in the same eight characters, so you can see they belong together. The zip holds your documents, so share it only with people who may read them.

You can revise a finished report. Changes of substance are recounted through the workflow. Every revision is kept.

Your documents stay with you

Your documents go to Anthropic under your Claude plan’s terms. Nothing goes to Causal Map. On Pro or Max, switch off “Help improve Claude” in Claude’s privacy settings before using client material.

Install

You need the Claude desktop app on a paid plan (Pro, Max, Team or Enterprise).

These are temporary instructions, until Rubicon is launched in Claude’s official plugin marketplace.

1. Open Settings from your account menu, then Plugins, Add, Add marketplace, and enter stevepowell99/causalmap-plugins.
2. Install Rubicon from it.

• Cowork (the main way). Start a Cowork chat with your documents folder attached.
• Code (optional). Use the Code tab, or Claude Code in a terminal. It needs Python and Node.js on your computer and offers to install them. In a terminal, type /plugin marketplace add stevepowell99/causalmap-plugins, then /plugin install rubicon@causalmap.

Try it first on this zipped sample folder of 18 interviews. The code is on GitHub, under the MIT licence.

Getting new versions

Check for updates yourself from time to time. In Settings, Plugins, click Add, then Manage marketplaces. Click the ⋮ beside causalmap-plugins and choose Check for updates.

Keeping track of your work

• Group your chats. Click the ⋮ on a Rubicon chat and choose Move to group. Then, in the view menu beside search, choose Group by > Custom groups.
• Ask in the folder. Each question Rubicon answers is a run with its own folder inside the documents folder, containing the report and all the details of how it was made. rubicon-runs.md lists them all. Each run’s folder also links back to the chat that made it. Attach the documents folder to a new Cowork chat and ask Rubicon what it has already answered.

What it costs

Nothing beyond your Claude subscription.

The plugin and an MCP server

Rubicon has a method layer and a trust layer. The table describes the next release, in which the Claude plugin and the MCP server share one method and one set of tools: the plugin runs the tools on your own computer, and the MCP server, once it is hosted, will offer them to other assistants such as Copilot. The plugin you can download today is the 9 October release, which works the older way.