This module covers how AI changes legal drafting and how to get the best results. Prompting is really the skill of giving clear instructions—something lawyers already do every day when they brief associates, define project scope, or draft engagement letters. The same discipline that makes you effective at those tasks makes you effective with AI.
How AI Changes Drafting
Traditional legal drafting starts from a blank page or a precedent document. You outline, write, revise, cite-check, and format.
AI-augmented drafting changes the starting point. Instead of a blank page, you begin with a structured output that you then review, revise, and make your own. The time savings on initial production are significant. But the professional responsibility for the final product remains entirely yours.
| Traditional | AI-Augmented | |
|---|---|---|
| Starting point | Blank page or precedent document | Structured AI output |
| Initial production | Hours of writing from scratch | Minutes of prompting |
| Where time goes | Writing, then revising | Prompting, then verifying and applying judgment |
| Error profile | Omissions, unclear reasoning, citation gaps | Fabricated citations, generic analysis, missed nuance |
Think of it this way: AI handles the assembly, and you handle the architecture, the judgment, and the quality control. Below is a walkthrough of how to make it happen.
The Outline Before the Draft
The most common mistake lawyers make with AI drafting is jumping straight to the end product. “Write me a motion to dismiss” will produce a generic output that requires heavy revision—if usable at all.
The problem is the same as handing an associate a stack of documents and saying “look into this.” You will get something back, but probably not what you needed.
A better option is to outline your argument structure first, then use AI to draft each section exactly as you envision it. You control the strategy. AI handles the production.
To supplement this outline, you can also give the AI your sources (e.g., cases, law, or past templates) and then ask it to draft. The more structure you provide upfront, the less revision you need on the back end. This is the same principle behind good legal writing—the real work happens before you start typing.
By using a thoughtful, detailed prompt and gathering the right sources, you can produce something you can work with in minutes. This is the difference between AI being useful and AI being a waste of time.
Six Core Strategies
Prompting techniques are principles that align with how language models process information, which we reviewed in Module 1. OpenAI’s and Anthropic’s prompting guides reach the same conclusions as to the best prompting methods.
Be explicit
State exactly what you want. Specify document type, jurisdiction, audience, tone, length, structure.
Provide context
Give the AI your role, your client's position, the procedural posture, and the relevant facts.
Break into steps
Divide complex drafts into sections. Draft each separately. Verify before building on prior output.
Ask it to reason
Instruct the AI to think step by step before drafting. This reduces errors on complex analysis.
Use examples
Show the AI a sample of the format, tone, or analysis style you want. It will mirror the pattern.
Permit uncertainty
Tell the AI it is better to flag doubts than to guess. This reduces fabrication.
Require Reasoning and Sources
As stated in the Research Module, one method to improve accuracy is to have AI produce the reasoning and the specific text that supports what it’s writing. In drafting, one helpful strategy is to have the AI generate citations that provide quotes from the actual source, which you can then verify.
Prompting Methods
Start zero-shot. Reach for few-shot or chain-of-thought when the task needs more structure.
Quick, straightforward tasks.
Zero-shot prompt
Summarize this NDA's key terms in plain English.
Model output
• Two-year term, auto-renewing. • Defined Confidential Information; carve-outs for independently developed or publicly available info. • Mutual return/destruction obligation on termination. • Delaware law; exclusive venue in New Castle County. • Liquidated damages capped at $50,000 per breach.
Below are some popular prompting methods. You can mix and match them to produce optimal results based on your needs.
| Method | What It Is | When to Use It |
|---|---|---|
| Zero-shot | A single instruction with no examples. | Straightforward tasks. Quick output you can refine. |
| Few-shot | Provide 2–3 examples of the output format before giving the real task. | Specific format or style. Consistent judgment calls across items. |
| Chain-of-thought | Instruct the AI to reason step by step before giving its answer. | Complex analysis. Multi-element legal tests. Anything requiring multiple reasoning steps. |
Most of the time, the best strategy is to start with zero-shot prompting and adjust from there.
Sometimes, you may need a more specific type of output. A complex drafting task might use few-shot examples to set the output format.
Or a chain-of-thought can work through the analysis and explicitly permit uncertainty to flag weak points.
Meta-Prompting
Meta-prompting as a back-and-forth
Instead of guessing a better prompt alone, ask the model what it needs and then answer it.
First prompt
Draft a motion to compel arbitration in our Jones matter.
Model critique
I can draft it, but the prompt is missing the court, governing agreement, procedural posture, anticipated opposition, page limit, and whether I should flag unsupported citations instead of guessing.
Refined prompt
Draft a motion to compel arbitration in Jones v. Acme, No. 24-cv-0187 (S.D.N.Y.). Governing agreement: Acme Master Services Agreement section 14 (attached). Posture: suit filed 2026-03-10; no responsive pleading yet. Anticipated opposition: Jones argues unconscionability and that the clause is unenforceable under New York law. Length: 15 pages including TOA. Structure: Statement, Standard, Argument (three sub-parts), Conclusion. Flag any citation you cannot verify.
Draft Zero output
Argument The Federal Arbitration Act reflects a strong federal policy favoring arbitration agreements according to their terms. Acme seeks an order compelling arbitration under section 14 of the Master Services Agreement, which requires binding arbitration of disputes arising out of the parties' commercial relationship. Jones does not dispute signing the agreement. The threshold question is therefore whether any generally applicable contract defense renders the clause unenforceable. On the present facts, the answer is no. [Verify New York authority on commercial unconscionability before filing.]
Meta-prompting works because the model first tells you what it needs, then you feed that context back into the real drafting prompt.
Another prompting technique is asking the AI to improve your prompt before you execute it. After writing your initial instructions, ask: “What additional information would help you give a better response to this prompt? What clarifications would improve the output?” Or try your own zero-shot prompt to get exactly what you need.
Some tools like Clio or Harvey already have this built in, but in our opinion, the best way is to come up with your own “prompt- improvement prompt.”
With this method, the AI will often identify missing context, ambiguities, or structural improvements you did not think of. This is often effective because the AI might otherwise have tried to compensate by hallucinating or producing a generally lower-quality product.
You can even go into a separate planning chat to refine your prompt. This loop—draft a prompt, ask for improvements, refine, execute—often produces better output than going straight to execution, especially on complex tasks.
Prompt refinement side by side
Use the planning chat to identify what's missing, then transfer those constraints into the revised prompt.
The technique also works in reverse. If you receive output that is not quite right, instead of rewriting the prompt from scratch, ask the AI: “How could I modify my original prompt to get output closer to what I need?” The AI can often diagnose why the output missed the mark and suggest specific prompt adjustments.
Draft Zero in Practice
Draft Zero to final
See what the lawyer adds: named facts, pinpoints, the governing test, and the right disposition rule.
Draft Zero
Plaintiff's claim must be dismissed because Defendant had no personal jurisdiction in this forum. Defendant is not located in New York and has no business there. The complaint should be dismissed. See Int'l Shoe Co. v. Washington, 326 U.S. 310 (1945); Daimler AG v. Bauman, 571 U.S. 117 (2014).
Generic, unsupported, citation names only. Structurally plausible, substantively weak.
Lawyer revision
The Court lacks personal jurisdiction over Defendant. Defendant is a Delaware corporation with its principal place of business in California; it maintains no offices, employees, or property in New York, and Plaintiff alleges no contacts arising out of or related to the Complaint's claims. Under Daimler AG v. Bauman, 571 U.S. 117, 137 (2014), a corporation is subject to general jurisdiction only where it is "at home" - ordinarily the state of incorporation or principal place of business. Because neither applies here, the Complaint must be dismissed under Rule 12(b)(2).
Pinpoint cite. Exact test. Named facts tied to the rule. Specific disposition under Rule 12(b)(2).
Module 2 introduced the concept of Draft Zero—the idea that AI output is somewhere between a blank page and a first draft. Here is what that looks like in a drafting workflow.
Outline
YouIdentify the argument structure, key authorities, and strategic choices.
Prompt
YouGive AI the outline, context, authorities, and specific instructions for each section.
Draft Zero
AIAI generates structured output based on your instructions.
Verify
YouCheck every citation and quotation against actual sources.
Revise
YouApply judgment. Sharpen the analysis. Adjust tone and strategy. Cut what does not serve the argument.
Final product
YouSign, file, or send the final document to a client.
The time savings come from steps 2 and 3. The professional value comes from steps 1, 4, 5, and 6—where the lawyer applies judgment to produce a polished product. AI compresses the production while the lawyer applies judgment.
Main Point
The same qualities that make a good brief—clear structure, precise language, specific authority, and logical reasoning—make a good prompt. With good prompts, thorough verification, and proper judgment, you can maintain or enhance work product quality while saving time.