How to Write Legal AI Prompts That Produce Usable Work in 2026

A prompt that returns a plausible but unusable answer costs more time than drafting by hand. The lawyer reads it, distrusts it, and starts again.

A legal AI prompt is an instruction that defines a legal task for a model. It names the task, the governing law, the source material, and the output format. That instruction sets the ceiling on the answer.

Prompting is no longer the only way legal teams direct AI. Work that once took a single instruction is increasingly saved as a named skill or handed to an agent. The instruction still governs what happens.

Five Parts of a Legal AI Prompt

A good AI prompt for lawyers specifies five things: role, intended outcome, context, output format, and constraints. Missing any one of them pushes work back onto the reader.

The same five parts apply whether the instruction goes to a general model, a legal platform, or an agent running several steps in sequence. Clarity is not a separate ingredient. It comes from naming format and constraints precisely.

Role and jurisdiction

Naming the role and the governing law stops a model defaulting to American framing on an English matter. One opening sentence does the work: “You are in-house counsel advising under the law of England and Wales.”

Jurisdiction belongs in the prompt even when it looks obvious from the document. A supplier agreement governed by Swedish law reads differently from the same terms under English law, and the model cannot infer which applies.

1. Intent: the outcome, not the topic

Intent tells the model what you want at the end, not what the document is about. “Summarise this NDA, focusing on jurisdiction and fees” produces a usable answer. “Summarise this document” produces a paraphrase.

Naming the outcome also sets the depth. A request to flag negotiation points returns something different from a request to explain the agreement to a non-lawyer. Both start from the same contract.

2. Context: the facts that change the answer

Context is the set of facts that change what the correct answer is. Contract type, counterparty relationship, commercial position, and applicable regulation all belong here. “In the context of a SaaS agreement” narrows a general answer into a relevant one.

Most of the time spent writing legal prompts goes on assembling that context by hand. Pasting clauses, naming the statute, restating the position. Connectors now remove much of that work, as covered further below.

3. Format: specifying what the output has to look like

Format decides whether the output is usable without reformatting. A risk table with named columns, a clause-by-clause list, a redline, or a plain-English summary under 200 words each demand different handling.

Unspecified format is the most common reason a technically correct answer still needs rewriting. Structure matters most in AI contract review, where a reviewer needs to act on findings rather than read prose.

4. Constraints: what the model must not do

AI is stopped from inventing case law by an explicit prohibition plus verification. Instruct it never to produce case names, citations, or statutes it cannot ground in the supplied material. Then check every reference before use.

Models do not apply that constraint on their own. Fabricated citations have reached court bundles in England and Wales, and solicitors have been referred to the regulator over them.

Constraints also cover scope. Tell the model to answer only from the attached documents and to say where they are silent. That beats an answer which quietly fills the gaps.

5. Refinement: the second prompt matters more than the first

Refinement is where most of the quality arrives. The first output shows what the model misread. The second prompt corrects it by narrowing scope, supplying a missing fact, or tightening the format.

A prompt that survives three rounds of refinement is worth saving rather than rewriting from memory next time. That saved version is the starting point for a skill.

Prompt Structures for Common Legal Tasks

The same handful of tasks account for most AI use inside a legal team. Writing each prompt from scratch wastes the refinement already done on it.

These skeletons are starting points to adapt, not fixed wording. Each one names a role, an outcome, a format, and a constraint.

The same structure carries across the ChatGPT prompts for lawyers used for research, drafting, and client correspondence. Litigation and compliance framings adapt the skeletons without changing the anatomy.

Keeping terminology consistent across a document set

Inconsistent terminology across a set of documents creates ambiguity a court may later have to resolve. Instructing the model to use “liability cap” throughout, or “indemnification” rather than “compensation”, fixes the immediate output.

Repeating that instruction in every prompt does not scale. Approved contract templates and legal playbooks solve it at source, because the agreed wording already sits in the document the model starts from.

Where Prompting Ends and Agents, Skills, and Plugins Begin

The difference between an AI prompt and an AI agent is scope of execution. A prompt returns one response to one instruction. An agent plans a sequence of steps, calls tools or systems, and produces a finished deliverable.

That distinction now shapes how legal teams deploy AI. Gartner projected that 40% of enterprise applications would include task-specific AI agents by 2026, up from under 5% in 2025.

Adoption inside legal has moved at a similar pace. Research published by the Association of Corporate Counsel found generative AI use in corporate law departments more than doubled in a year. It rose from 23% to 52% among US respondents.

None of this displaces prompting. Agents still act on instructions, and legal AI tools still return weak output from vague ones. What changes is where the instruction lives and how often anyone rewrites it.

What an agent does that a prompt cannot

An agent handles work spanning several steps and more than one system. Extracting obligations across a supplier portfolio, checking each against policy, and routing the exceptions for review is agent work. Summarising one agreement is not.

The gain is that the sequence runs without a person prompting each stage. The cost is that more completes before anyone reads the result, which is why logged actions and permission limits matter more here.

Autonomous legal research remains the weakest application. An agent that searches, cites, and drafts with no verification step is the fastest route to a fabricated authority reaching a filing.

Skills and plugins: prompts saved as repeatable workflows

AI skills and plugins are prompt workflows saved under a name so anyone on the team can run them the same way. Invoking a review or triage workflow replaces writing the instruction again from memory.

Practice-area plugin suites now ship for legal work, packaging contract review, NDA triage, compliance checks, and templated responses as named commands. The prompt engineering happens once, at authoring time.

The consequence for a legal team is standardisation. One well-drafted prompt becomes the floor for everyone rather than the private habit of whoever wrote it, and each improvement applies to every future run.

Connectors: giving the model the contracts instead of pasting them

Connectors remove the copy-and-paste step most legal prompting still depends on. The Model Context Protocol is an open standard that lets AI tools connect securely to external systems. The model retrieves source material under access controls rather than waiting to be handed it.

Context stops being something the lawyer assembles and becomes something the system supplies. A question about a named counterparty’s NDA no longer requires finding the file, opening it, and pasting the relevant clauses.

Contract platforms have started exposing their repositories through that standard. Miramis Connect puts Miramis contract management inside Claude, ChatGPT, Copilot, and Gemini using OAuth 2.1 authentication.

The scope reaches past retrieval. Drafting from an approved template, routing a contract for approval, and sending it for signature all run from the same conversation.

Permission scoping is what makes that workable. Access follows the rights each person already holds in the platform, and customer data is not used to train or fine-tune AI models.

Choosing Between a Prompt, a Skill, and an Agent

A legal team should use an agent instead of a prompt when the work repeats, spans several steps, and touches a system of record. One-off analysis of a single document stays faster as a prompt.

Most legal AI use still starts as a prompt. The progression matters more than the label. A prompt rewritten weekly should become a skill, and a skill that keeps triggering the same follow-up actions is a candidate for an agent.

Verification and Confidentiality

Accountability does not move to the tool. A solicitor stays responsible for the work regardless of what produced the first draft, a position set out in Law Society guidance on generative AI.

Building verification into the workflow

Verification means checking every citation, figure, and clause reference before output reaches a client, a counterparty, a filing, or a contract draft. The check belongs inside the workflow, not in whatever time is left afterwards.

Agent workflows raise the stakes, because more steps complete before a person sees the result. Audit trails and logged actions are what make a multi-step output reviewable rather than something accepted on trust.

Scope limits set at the prompt stage reduce how much verification is needed later, which is much of what ethical prompting for legal work addresses.

What should not go into a general-purpose model

Lawyers should not enter privileged or client-identifiable information into consumer-tier general-purpose models, where retention and training terms sit outside the firm’s control. Non-confidential drafting, explanation, and formatting tasks carry far less exposure.

Enterprise deployments and permission-scoped connectors sit in a different position. Access is controlled, actions are logged, and data handling is contractual, which makes security and compliance review the deciding factor rather than the model itself.

Conclusion

Prompts, skills, and agents have become essential tools for legal teams working with AI. Each technique serves a specific purpose; prompts for single tasks, skills for repeatable workflows, and agents for multi-step execution. Well-formed instructions remain the foundation regardless of which approach a team uses.

However, while general-purpose AI tools continue to evolve, specialised platforms designed explicitly for legal tasks—such as Miramis contract lifecycle management—offer something general models cannot provide: governed context. That includes the executed agreements, approved templates, playbook positions, and approval rules that decide what gets signed and by whom.

By choosing contract lifecycle management infrastructure built for the full lifecycle, legal teams can maximise accuracy, maintain oversight, and scale contract work without scaling headcount. The platform holds the context. Prompts, skills, and agents are how people and systems act on it.