---
title: "Install the Judgment agent skill"
description: "Install Judgment tracing, evaluation, and SDK guidance into a coding agent for one repository and verify that the agent loads it."
sidebar:
  label: "Install the Judgment agent skill"
seo:
  title: "Install the Judgment Agent Skill"
  description: "Add the Judgment skill to a repository for Cursor, Claude Code, Windsurf, or Codex."
---

Install the Judgment skill in the repository where your coding agent will work.
The skill teaches the agent to fetch current Judgment documentation, instrument
the application's real execution path, evaluate outputs, test agent changes
offline, choose scorer patterns, and build Python code judges with supported SDK
practices.

> **Info**
>
> A skill is instruction, not access. It does not install the Judgment SDK,
> connect an MCP server, or grant credentials. [Connect the data
> MCP](/documentation/mcp-and-agent-tools) when the agent needs project data.

## Prerequisites

- Node.js and `npx` on your path.
- Cursor, Claude Code, Windsurf, Codex, or another Agent Skills-compatible
  coding agent.
- The repository where you want the instructions to travel with the code.

## 1. Install from the repository root

```bash
npx skills add JudgmentLabs/skills --skill judgment
```

Choose the coding agents used by the project when the installer prompts. A
project-local install keeps one canonical skill directory and wires it into the
selected agents' discovery directories. The exact link paths vary by client;
keep the canonical skill and generated links together when committing them.

For non-interactive setup, select the supported agent names and accept the
install explicitly. For example:

```bash
npx skills add JudgmentLabs/skills \
  --skill judgment \
  --agent codex \
  --agent claude-code \
  --yes
```

Run the command again with the agent names your team actually uses rather than
copying unsupported directories by hand.

## 2. Verify discovery

Review the destinations printed by the installer. Each selected agent should
resolve a `SKILL.md` for the Judgment skill, whether the destination is the
canonical directory or a symlink to it.

Then start a new agent session in the repository and ask:

```txt
Load the Judgment skill. Without changing code, summarize its recommended first
steps for adding Judgment to this application, including when to start with
tracing and how to choose between hosted scorers and Python code judges.
```

Success means the agent identifies the installed skill, starts from current
documentation and the application's real path, recommends tracing unless the
task specifically asks for evaluation or code judges, and distinguishes
rubric-based hosted scorers from deterministic or dependency-heavy Python code
judges. If it cannot find the skill, restart the agent and check the
installer-reported path from the repository root.

## 3. Use the skill on a code task

Give the agent a bounded implementation or review task. For example:

```txt
Use the Judgment skill to audit tracing in the checkout agent. Fetch the current
Judgment tracing guidance, follow the execution path that handles one checkout,
and identify missing spans, attributes, or flush behavior. Propose the smallest
supported SDK change, but do not edit files yet.
```

Connect the [Judgment data MCP](/documentation/mcp-and-agent-tools) separately
when the task also needs live project data. See [Use Judgment from a coding
agent](/documentation/mcp-and-agent-tools/coding-agents) to combine the skill,
Judgment MCP server, and repository context without mixing their roles.