Intellectual Property Clause Isolation Prompt

This prompt is built for legal engineering teams who need to extract structured intellectual property (IP) clauses from contracts, technology transaction agreements, and collaboration documents with precise citation boundaries. The job-to-be-done is isolating IP ownership, license grants, assignment rights, work-for-hire provisions, and joint development IP allocation into a machine-readable schema that downstream systems can use for risk scoring, obligation tracking, or review queue routing. The ideal user is an engineer or legal operations professional building a contract intelligence pipeline where IP allocation errors carry material business risk—missed assignment clauses, hallucinated license scopes, or conflated background and foreground IP can lead to incorrect ownership assertions in production systems.

Use this prompt when your pipeline requires clause-level extraction with exact section, page, or paragraph citations. It is designed for high-stakes workflows where hallucinated clause content is unacceptable and where every extracted field must be traceable back to source text for human verification. The prompt expects a clean text representation of a contract or agreement as input, along with an output schema that defines the IP fields you need. It works best when paired with a validation layer that checks citation boundaries, confirms extracted text matches the source, and flags low-confidence extractions for human review. Do not use this prompt for general contract summarization, end-user legal advice, or workflows where approximate answers are acceptable.

Before deploying, ensure you have defined your IP taxonomy clearly—what counts as a license grant versus a covenant not to sue, how you distinguish assignment from exclusive license, and what joint development IP allocation looks like in your domain. The prompt's effectiveness depends on this upfront schema design. Wire it into a pipeline that includes pre-processing for document structure preservation, post-extraction validation against source text, and a human review queue for clauses flagged with low confidence or missing citations. Avoid using this prompt on heavily redacted documents, poor OCR output, or agreements where IP terms are scattered across multiple incorporated-by-reference documents without first resolving those references.

This prompt template isolates intellectual property clauses from a contract or agreement document. It is designed for legal engineering teams building contract review systems where clause-level extraction with precise citation boundaries reduces hallucination risk. The prompt requires two inputs: the full document text and a target output schema. It instructs the model to extract verbatim clause text, classify each clause into a controlled IP taxonomy, distinguish background from foreground IP, and flag missing provisions explicitly rather than inventing content.

textCopy
Extract intellectual property clauses from [DOCUMENT_TEXT] below. Return only the structured JSON output specified in [OUTPUT_SCHEMA]. Do not add commentary, analysis, or legal advice.

## Instructions
1. Identify every clause, section, or provision that addresses intellectual property ownership, licensing, assignment, or work-for-hire.
2. For each identified clause, extract the full verbatim text with exact section, page, and paragraph citations.
3. Classify each clause into one or more of these categories: IP_OWNERSHIP, LICENSE_GRANT, ASSIGNMENT, WORK_FOR_HIRE, JOINT_DEVELOPMENT, BACKGROUND_IP, FOREGROUND_IP, MORAL_RIGHTS, THIRD_PARTY_IP, IP_REPRESENTATIONS, IP_INDEMNIFICATION, RESIDUAL_RIGHTS.
4. Distinguish between background IP (pre-existing, owned before the agreement) and foreground IP (developed during the agreement term).
5. For license grants, extract scope boundaries: exclusive vs. non-exclusive, territory, field of use, sublicensing rights, and duration.
6. For assignment provisions, identify trigger conditions, consent requirements, and whether assignment includes future IP.
7. For work-for-hire provisions, identify the commissioning party, the creator party, and whether the provision explicitly uses "work made for hire" language.
8. Flag joint development IP allocation rules, including default ownership splits and decision-making authority for jointly created IP.
9. Set confidence to LOW if the clause language is ambiguous about IP allocation, uses undefined terms, or cross-references missing sections.
10. Set is_absent to true for each expected IP category that does not appear in the document.

## Constraints
- Never invent clause text. If text is not present, set clause_text to null.
- Never interpret silence as a default rule. Flag missing provisions explicitly.
- Cite exact section numbers, page numbers, and paragraph numbers where available.
- If the document contains no IP provisions at all, return a single entry with category NO_IP_CLAUSES and is_absent set to true for all expected categories.

## Document Text
[DOCUMENT_TEXT]

## Output Schema
[OUTPUT_SCHEMA]

To adapt this prompt for your pipeline, replace [DOCUMENT_TEXT] with the full contract text after preprocessing for layout and OCR artifacts. Replace [OUTPUT_SCHEMA] with your target JSON schema, which should include fields for clause_text, category, citation, confidence, scope_details, and is_absent. If your document source includes page numbers and section headings, preserve those in the input text so the model can produce accurate citations. For multi-document workflows, run this prompt once per agreement and merge results with document-level metadata. Always validate the output against your schema before ingestion—malformed JSON or hallucinated clause text are the most common failure modes in production.

Before deploying, build an eval harness that checks citation boundary accuracy, category classification precision, and null-handling correctness. Test with documents that contain no IP clauses, ambiguous language, and cross-references to missing sections. If your use case involves regulated agreements or high-value transactions, route LOW-confidence outputs to human review and log all extractions with document version and model version for auditability. Do not treat this prompt as legal advice—it is an extraction tool that requires downstream verification.

The Intellectual Property Clause Isolation Prompt is designed to operate as a deterministic extraction step within a larger document processing pipeline, not as a standalone chat interaction. To achieve production-grade reliability, the prompt must be wrapped in a harness that handles input normalization, output validation, retry recovery, and human escalation. The goal is to ensure that every extracted clause is traceable to source text and that missing or low-confidence extractions are surfaced before they propagate into downstream systems like obligation trackers or risk dashboards.

Pre-processing is the first gate. Before the prompt ever sees the document, strip headers, footers, and page numbers that interfere with section numbering. Normalize whitespace but preserve paragraph breaks. If the document is a scanned PDF, run OCR and validate text quality—garbled OCR output will produce hallucinated clause text and false negatives on critical IP provisions. Map party names to a normalized entity list before extraction so that 'Licensor,' 'Company,' and 'Acme Inc.' are consistently attributed across clauses. This normalization prevents the model from treating the same party as two different entities in the output JSON.

At the API layer, call this prompt via chat completions with response_format set to json_object if the model supports it. Set temperature to 0 for deterministic extraction. Use a system message that constrains the model's role: 'You are a contract clause extraction system. Return only valid JSON. Do not provide legal analysis or commentary.' This prevents the model from drifting into explanatory prose or legal advice, which would break downstream JSON parsers. After extraction, validate the JSON output against the schema. For every clause_text field, perform substring matching or fuzzy matching with a minimum similarity threshold of 0.95 against the source document. Any clause_text that does not match source text with high confidence must be flagged for human review—this is the primary defense against hallucinated clause content.

Retry logic handles transient failures without infinite loops. If validation fails due to malformed JSON, retry once with the error message appended to the prompt so the model can self-correct. If validation fails due to missing required fields, retry with explicit instruction to populate those fields or set them to null. Stop after two retries and escalate to human review. For routing, send extractions with confidence below [CONFIDENCE_THRESHOLD] to a review queue. Route extractions where is_absent is true for IP_OWNERSHIP or LICENSE_GRANT to a senior reviewer—missing core IP provisions in a technology agreement may indicate an incomplete document or a significant gap that requires immediate attention before the contract proceeds.

Logging is non-negotiable for auditability. Log the prompt version, model version, document hash, extraction timestamp, confidence scores, and validation pass/fail status for every extraction. This audit trail enables retrospective analysis when extractions are questioned and supports prompt iteration by correlating failure patterns with specific document characteristics. Before deploying, run the harness against a golden dataset of contracts with known IP clause locations and verify that the pipeline correctly extracts, cites, and flags every provision. Do not ship this prompt without the validation layer and human review gate in place—unguarded extraction in IP-heavy agreements creates legal risk that no prompt alone can mitigate.