NDA Triage

NDA Triage is an official Anthropic skill from the knowledge-work-plugins repository. It helps teams screen incoming non-disclosure agreements consistently before a signature is considered. The skill is not legal advice and does not replace qualified counsel or the legal department responsible for the organization. According to the provider, its purpose is to classify an NDA quickly and route it to the appropriate expert review path. The three intended outcomes are GREEN for standard approval, YELLOW for counsel review, and RED for full legal review or negotiation.

Workflow and input material

According to the provider, the skill accepts an uploaded PDF or DOCX document, a link to a document system, or pasted agreement text. If no NDA is supplied, the agent should ask the user to provide one. Before screening, the skill looks for a local NDA playbook, such as an organization-specific file defining requirements for mutuality, terms, carveouts, and prohibited provisions. If no playbook is configured, it proceeds with clearly identified market-standard defaults. Those defaults include mutual obligations, a bounded term, customary exclusions for information already known, public information, independent development, and lawful third-party receipt, plus disclosure when required by law or legal process.

First-pass review framework

NDA Triage structures its first review around concrete agreement characteristics. It identifies the agreement structure and distinguishes mutual from unilateral arrangements. It then examines the definition of confidential information, marking requirements, and standard exclusions. Next it considers the care standard, purpose limitation, and permitted recipients such as employees, contractors, advisors, professional consultants, and affiliates. The workflow also addresses agreement term, survival of confidentiality, return and destruction, lawful retention, equitable relief, and any predetermined damages.

The skill gives special attention to provisions that, according to the provider, do not belong in an ordinary NDA. These include employee non-solicitation, non-compete, exclusivity, standstill obligations, broad residuals language, hidden intellectual-property assignments or licenses, unusual audit rights, and an unreasonable governing-law, jurisdiction, or arbitration arrangement. It can also flag when a document called an NDA contains substantive commercial obligations or is only a confidentiality section inside a broader agreement. That distinction helps prevent a short confidentiality screen from being mistaken for a complete contract review.

Classification and report

GREEN is reserved, according to the provider, for agreements whose structure, carveouts, terms, disclosure permissions, return provisions, and remedies fit the standard. YELLOW identifies limited deviations that counsel may be able to resolve in a focused review. RED identifies significant issues such as missing critical exclusions, an embedded non-compete, an unreasonable or perpetual obligation, a hidden license, liquidated damages, or a document that is really another type of contract. The generated report is intended to state the parties, agreement type, term, governing law, review basis, individual screening results, specific issues, risk, suggested changes, and next steps.

Boundaries, security, and practical use

An NDA may contain trade secrets, personal data, and sensitive negotiation context. Users should therefore use only approved clients and storage locations, limit access, and determine before upload what information may be transmitted to a model or connected service. The skill cannot decide enforceability, jurisdiction, or the organization’s acceptable risk on its own. Even a GREEN result must be used within internal delegation rules. According to the provider, uncertain issues and especially YELLOW or RED results require review by qualified legal professionals. The official repository publishes this material under the Apache-2.0 license, subject to that license’s conditions. GitHub reports exactly 23,935 stars for the repository as a whole on September 9, 2026; that point-in-time number is a popularity signal for the collection, not evidence of legal quality or security. The individual skill directory has no separate star count.