Value Alignment Toolkit for Cowork

Cowork & skills

Three words carry all the technical weight on this site - Cowork, plugin, and skill. This page explains each one in plain language, with no assumed background.

What is Cowork?

Cowork is a mode inside Anthropic's Claude desktop application. In ordinary chat mode, the AI assistant only talks with you. In Cowork mode, it can also work directly with files on your computer and with other tools you've connected - reading a document you point it to, drafting a new file, filling in a table, working through several steps in a row without you having to prompt each one individually.

A simple way to picture it: a capable assistant who sits down at your computer, works through the parts of a task that are mechanical - reading, organizing, drafting - and then hands you back a draft to check. You stay in charge of the judgment calls; it carries the legwork.

It helps to be clear about where this technology stands. Cowork is currently what Anthropic calls an early "research preview" - plain English, a work in progress that is still being tested and changed, not a finished, stable product. Anthropic is actively adjusting how it looks and behaves. That doesn't make it unusable, but it does mean this page describes the general shape of how it works rather than a fixed set of screens that will look the same next month.

What is a plugin, and what is a skill?

A skill is a packaged set of instructions that teaches the assistant how to do one specific task well - for example, "design a legal department metrics framework, the way TRI's guide describes it." A skill tells the assistant what questions to ask, what order to ask them in, and what the finished draft should contain.

A plugin is a bundle you install that delivers a group of related skills together, along with any reference material those skills need to do their job well. The Value Alignment Toolkit plugin is built to bundle skills from TRI's own guides - one for each artifact a legal department currently produces by hand. So far only the first, Design metrics, is actually built; it's the working example being shared here (see The skills for the full list).

In plain terms: think of the plugin as a toolbox and each skill as one specialized tool inside it. You install the toolbox once; the assistant picks the tool that fits what you ask for.

One detail explains why the skills stay faithful to TRI's method instead of guessing at it: the plugin carries TRI's own source material - the guides, the Metrics Library, the worksheets - bundled inside it in a clearly labeled reference folder, separate from the mechanical parts of the plugin. The skills were built by reading that material, so they follow TRI's actual method and terminology rather than a generic approximation of it.

How do you install it, and how does a session work?

The general shape is simple, even though the exact screens involved will change as Cowork itself changes. You install the Value Alignment Toolkit plugin into Cowork once. After that, using a skill doesn't require you to open a menu and pick one by name - you just start a conversation and describe what you want in your own words, something like "help me build our legal department's metrics framework." The assistant recognizes which skill fits what you've asked for and follows it, asking you for documents or answers where it needs them, and handing back a draft when it has enough to work with.

Where this might be wrong - please check: this page deliberately does not describe exact menu names, button labels, or a click-by-click installation sequence, because the Cowork interface is an early research preview and is actively changing. Before this page is published, its description of installation and session flow should be checked against current Cowork documentation and updated with the actual steps.

The basic workflow, end to end

Put together, a working session with the plugin generally follows the same shape from start to finish:

1. Install once

Add the Value Alignment Toolkit plugin to Cowork. This is a one-time setup step, not something you repeat per task.

2. Start a conversation

Describe what you're trying to produce, in plain language - no need to name a skill or know which one applies.

3. Provide documents or answers

The assistant asks TRI's own questions and may ask you to share a document, such as your current metrics or a past board slide.

4. Receive a labeled first draft

You get back a draft artifact - a table, a set of definitions, a slide - with any gaps or weak points clearly flagged, not hidden.

5. Review, correct, and finish

You check the draft against your own judgment, fix or override anything that needs it, and decide what actually goes forward.

You make the final call. Every output is a labeled first draft, and the legal and business judgment stays with you throughout.