How to use the AI Tool Handoff for Legal Review agent skill

What is an agent skill and what does this one do?

An agent skill is a folder containing a SKILL.md file with instructions that an AI assistant reads when a request matches the skill's description. This skill tells the assistant how to work with dedicated bulk-review software in legal due diligence.

The skill is built on one split of work. Luminance and Kira are good at reading hundreds of contracts and finding a clause type such as change of control. They are weaker at judgment, such as deciding whether a change-of-control provision is actually triggered by the deal structure. The skill hands the bulk extraction to the tool, then runs a QA layer on what comes back, then applies judgment and drafts a summary in the house format. It does not run the tool itself.

• Checks the firm's CLAUDE.md config for the tool in use, the clause types to extract and the trust level
• Prepares a load request for a person or team to run in the tool
• Samples the tool's flagged documents and compares them to the actual clauses, according to the trust level
• Applies judgment to each flagged provision, such as stock sale versus asset sale or how the contract defines change of control
• Produces a QA summary with counts for flagged, triggered and material items

Check whether the skill fits your batch

The source is clear that this skill is for large, uniform batches. Hand off only when all of the following are true: the category has more than 50 documents, the clause type is one the tool handles well (such as change of control, assignment, exclusivity, MFN, termination or auto-renewal), and the documents are reasonably uniform, such as customer contracts on similar paper.

Do not use it for bespoke or heavily negotiated documents, side letters or amendments, or any question about what a clause means for the deal rather than whether it exists. For a few hundred documents with a defined column schema, the source says native tabular review (/corporate-legal:tabular-review) is faster to set up and keeps work product local.

Install the skill from ZIP or the Claude apps

Option one: download the ZIP from the directory page and unzip it into your user skills folder, ~/.claude/skills/, or into a project's .claude/skills/ folder. The unzipped folder should contain SKILL.md.

Option two: upload the ZIP in the Claude apps under Settings → Capabilities → Skills.

• User-wide install: ~/.claude/skills/
• Project install: the .claude/skills/ folder inside your project

Set up the required CLAUDE.md config

The skill reads ~/.claude/plugins/config/claude-for-legal/corporate-legal/CLAUDE.md. Its AI-assisted review section must name the tool in use (Luminance, Kira or none), the clause types it is used for, the trust level and the handoff process, meaning who loads documents and who performs QA.

If that file says no AI tool is in use, the skill does nothing and everything goes through diligence-issue-extraction directly. The skill also expects the companion corporate-legal tabular-review and diligence-issue-extraction skills to be referenced by that file.

• Trust level options named in the source: use as-is, spot-check X%, or full human review of flagged documents
• The trust level is set in CLAUDE.md, usually at cold-start, based on the team's experience with the tool

Trigger the skill with natural requests

You do not need a special command. The skill's description lists phrases that should activate it, and the assistant also uses it when diligence-issue-extraction hits a high-volume category. Plain requests that match the description work best, and naming the deal, category and tool helps.

• "Send to Luminance" for the change-of-control review on the Acme deal
• "Bulk review the 300 customer contracts for assignment clauses"
• "Run AI extraction on the supplier agreements and QA the output"
• "Prepare a Kira load request for the exclusivity provisions in the data room"

What the output looks like

The skill produces a load request when someone else runs the tool, and a summary when results come back. The summary records the tool used, documents processed, extraction targets, trust level, sample size, error rate and whether the sample was accepted, widened or sent to full re-review. It then shows a results table comparing documents flagged, documents actually triggered after judgment, and material items.

The source states that the summary is derived from privileged or confidential VDR documents. It should be stored with the matter's privileged files, because sharing it beyond the privilege circle can waive privilege. The skill ends with a next-steps decision tree that you choose from.

• Load request: document count, VDR folder path, load destination, extraction targets, and a note to return only flagged results
• QA summary: trust level, sample size, error rate, results table, and counts of findings and consents added to the closing checklist

Safety notes: scripts and network

According to the source, this skill runs no scripts and uses no network access. It works by reading your CLAUDE.md config and the documents you provide, and writing its outputs to the matter folder. Matter-specific files are read only from the active matter, unless the Cross-matter context setting is on.

The skill is not a substitute for the tool's interface. A human user, or the tool's own interface, must run the extraction. Do not rely on tool output for complex or mixed document sets without human review, because the source warns that missed issues are a real risk.

• Outputs are written to ~/.claude/plugins/config/claude-for-legal/corporate-legal/matters/<matter-slug>/ when matter context is enabled
• Keep outputs with privileged files