Legal Automation Process Explained

Baoyu on X: "1. Skills are skills, domain knowledge, workflows, etc., equivalent to a manual on how to excel at something.

For example, in the repository at github.com/anthropics/cla, there is a skill called nda-review under the commercial-legal/skills/ folder. Inside is a SKILL.md file that clearly outlines which clauses to compare first when reviewing an NDA, follows the team's playbook at https://t.co/597jwLhM8O" / X

1. Skills are skills, domain knowledge, workflows, etc., equivalent to a manual on how to excel at something. For instance, in the github.com/anthropics/cla repository, there is a skill named nda-review located in the commercial-legal/skills/ directory. It contains a SKILL.md file detailing which clauses to compare first during an NDA review, following the team's playbook, grading contracts with green, yellow, or red flags, identifying situations requiring escalation, and specifying the output format as Word revision mode. This serves as a work manual for Claude, but it does not perform tasks itself.

1. An Agent is the entity that actually executes tasks. Besides the main executing Agent, custom Agents typically fall into two categories: Subagents and Scheduled agents.

2.1 Subagents are "clones" dispatched to handle specific tasks independently. For example, in the repository, the corporate-legal:tabular-review skill requires tabulating hundreds of contracts in a data room. If the main conversation were to read each document sequentially, the context would quickly exceed limits. Therefore, it dispatches subagents, with each subagent handling one document in parallel, and consolidates results back to the main conversation. The main Agent only sees the final table, while the intermediate information from hundreds of reads is isolated.

2.2 Scheduled agents are background tasks that run automatically at set intervals. For instance, renewal-watcher scans the contract library weekly, listing contracts expiring within 90 days and posting them to a designated Slack channel. You don't need to remember dates; it handles that for you. Similarly, docket-watcher (monitors court case updates) and reg-feed-watcher (tracks new regulatory rules) operate this way.

1. MCP connectors are connectors that bring external data into the system. No matter how well-written the skills are, they need contracts to review. The repository includes MCP connectors like Ironclad (contract repository), DocuSign (signed contracts), and iManage (document management). Agents use these MCP connectors to access real company contract repositories rather than manually copying and pasting.

Similarly, the litigation plugin connects to Everlaw (e-discovery), CourtListener (federal court rulings database), and Trellis (state court database). Changing practice areas involves switching to a different set of data connectors.

1. Plugins are containers that bundle all these components together. The commercial-legal plugin folder contains:

• A collection of skills (nda-review, vendor-agreement-review, escalation-flagger, etc.)
• Several scheduled agents (renewal-watcher, deal-debrief)
• A .mcp.json file that tells Claude which external systems to connect to
• A CLAUDE.md template for recording your team's playbook

Installing this single plugin equips you with a complete suite of enterprise contract review capabilities.

Quote

changbo

@mrnetman

1h

Replying to @dotey

Could the expert explain what Claude is trying to achieve by using plugins, skills, and agents?