- GLM Legal prompts separate supplied facts, task goals, and jurisdiction before requesting analysis
- Citation rules require every finding to reference the source clause or document
- Output schemas make responses structured, stable, and reviewable by legal professionals
- Quality controls distinguish facts from analysis and flag missing information
- Human review remains mandatory before any output is used in practice
What Makes GLM Legal Prompts Different
A reliable legal prompt does more than ask a question. It separates the factual record from the requested task, defines the legal context, and specifies exactly how the answer should be structured. GLM Legal prompts follow this discipline because legal work demands traceable, verifiable output rather than conversational guesses.
The core principle: define facts, legal issues, applicable law, output structure, and review requirements explicitly in the prompt. Vague prompts produce vague analysis; structured prompts produce work products a lawyer can actually check.
Background Facts
- Contract type, parties, objective
- Applicable law and jurisdiction
- Review priority context
Task Definition
- Specific legal objective
- Numbered deliverables
- Explicit inference limits
Citation Rules
- Clause or page references
- No unsourced facts introduced
- Confidence levels required
Output Schema
- Fixed column structure
- Separate facts from analysis
- Escalation fields included
Never let the model infer the jurisdiction or the source set. A prompt that states the governing law, the exact documents supplied, and the required citation format will consistently outperform an open-ended request.
Core Prompt Patterns by Legal Task
Different legal workflows need different prompt structures. The table below maps the three primary GLM Legal prompt patterns to their intended use.
| Prompt Pattern | Task Goal | Input Materials | Output Focus |
|---|---|---|---|
| Contract review | Identify obligations, unusual clauses, risks | Contract text, exhibits, playbooks | Issue list, risk classification, follow-up questions |
| Evidence organization | Build chronology and issue map | Chronologies, emails, pleadings | Fact matrix, timeline, contradiction list |
| Compliance checklist | Map requirements to controls | Regulations, policies, risk registers | Requirement matrix, gaps, responsible roles |
Contract review prompt structure. State the contract type, parties, business objective, jurisdiction, and review priority. Then ask for clause references, issue categories, plain-English explanations, business impact, suggested revisions, and confidence levels — one numbered item per finding.
Evidence organization prompt structure. Define the matter, parties, jurisdiction, and issues under review. Require a dated chronology with source documents for each event, distinguish observed facts from allegations, and prohibit resolving contradictions without explaining the competing evidence.
Compliance checklist prompt structure. Scope the organization, activity, jurisdiction, regulation version, and review period. Map each requirement to a summary, business process, evidence, responsible role, frequency, status, and remediation action.
Always specify the governing jurisdiction and the regulation version in the prompt. If applicability cannot be established from the supplied materials, the prompt should instruct the model to mark the item for human verification rather than guess.
Building Your First GLM Legal Prompt: Step by Step
State the background facts
List contract type, parties, business objective, applicable law, and review priority as bullet points. Keep facts separate from instructions so the model cannot conflate the two.
Define the task with numbered deliverables
Ask for specific outputs — for example, clause reference, issue category, explanation, impact, follow-up action, and confidence — numbered so each appears in every finding.
Add citation requirements
Require a source citation for every material conclusion. Explicitly forbid introducing facts not present in the supplied documents.
Specify the output schema
Define exact columns or fields: Issue ID, Clause reference, Risk category, Finding, Impact, Follow-up action, Confidence. A fixed schema makes outputs comparable across runs.
Insert quality controls
Add instructions to restate the task, separate facts from assumptions, identify missing information, and avoid presenting draft analysis as a final legal conclusion.
Run the prompt on a small, known document set where you already know the answers. Verify citation behavior and factuality before deploying it on production legal materials.
Quality Controls and Common Prompt Failures
Quality controls are the difference between a demo and a dependable workflow. The prompt-quality-controls pattern below can be prepended to any legal prompt.
| Control | What It Does | Failure It Prevents |
|---|---|---|
| Restate the task | Forces the model to confirm understanding | Drifting off-objective answers |
| List facts vs assumptions | Separates supplied record from inference | Fabricated factual premises |
| Identify jurisdiction | Anchors analysis to governing law | Wrong-jurisdiction reasoning |
| Cite every conclusion | Ties findings to sources | Unsupported legal claims |
| Flag missing information | Surfaces gaps that change results | Silent, confident errors |
| Fixed output schema | Locks response structure | Unparseable, inconsistent output |
Recommended control set for production prompts:
- Define the jurisdiction explicitly
- Specify the source document set
- Require clause or page citations
- Separate facts from analysis
- Use a fixed output schema
- Include an escalation or review field
Legal language models can produce fluent but incorrect citations and quotations. Every legal authority, quoted passage, date, and party name in the output must be checked against the original sources before use.
Output Schemas and Review Workflow
A well-designed output schema turns a model response into an auditable work product. Pair each schema field with a human review responsibility.
| Schema Field | Example (Contract Review) | Human Review Responsibility |
|---|---|---|
| Issue ID | C-003 | Track through resolution |
| Clause reference | Section 4.2 | Verify quoted text |
| Risk category | Commercial risk | Confirm business context |
| Finding | Auto-renewal without notice period | Check interpretation |
| Follow-up action | Request 90-day notice clause | Negotiate revision |
| Confidence | Medium (supplied text only) | Weight accordingly |
Prompt Deployment Checklist:
- Background facts separated from task instructions
- Jurisdiction and applicable law stated explicitly
- Citation requirement included in every prompt
- Fixed output schema defined and tested
- Human review step documented before production use
FAQ
Q: What is the basic structure of a GLM Legal prompt?
A GLM Legal prompt contains four blocks: background facts (parties, contract type, jurisdiction), a numbered task definition, citation requirements, and a fixed output schema. Separating facts from instructions keeps the model anchored to the supplied record.
Q: Why do I need to specify the jurisdiction in every prompt?
Legal analysis depends on governing law. Without an explicit jurisdiction, the model may reason under the wrong legal framework or mix rules across jurisdictions. Stating the applicable law and regulation version is a required quality control.
Q: Can GLM Legal prompts replace lawyer review?
No. Model output is work-product assistance, not a final legal conclusion. A qualified lawyer should verify citations, confirm applicability, and approve any material used in practice. Prompt quality controls reduce errors but do not remove the review requirement.
Q: How do I stop the model from inventing facts?
Add explicit rules: forbid introducing facts not present in the supplied documents, require citations for every conclusion, and instruct the model to flag missing information rather than fill gaps with assumptions.