How to Write AI Prompts
That Actually Work for Your Law Firm

We almost published a blog post last year that said Texas follows a “pure comparative negligence” standard. It doesn’t. Texas uses modified comparative negligence with a 51% bar — if you’re 51% or more at fault, you recover nothing. ChatGPT wrote that draft. Nobody on the content team caught it. The attorney reviewing it before publication did, because that’s what attorneys do, but it was a wake-up call about how we were using AI in our workflow.

The post wasn’t bad because AI wrote it. It was bad because the prompt that generated it was lazy. “Write a blog post about comparative negligence in car accident cases.” That was the entire instruction. No jurisdiction. No audience. No constraints on what to verify. We handed a tool trained on the entire internet a vague request and got a vague, partly wrong answer back.

That experience changed how we use AI tools for every law firm client we work with. The rest of this guide is what we learned — not from reading about prompt engineering, but from screwing it up on real client work and fixing the process. If you’re using generative AI in your firm’s marketing and the output keeps disappointing you, the problem is almost certainly in the instructions, not the tool. Our AI search guide for lawyers covers the broader landscape, but this is the hands-on piece.

Legal Content Is Where Bad Prompts Hurt the Most

Here’s why this matters more for your firm than for, say, a SaaS company writing about project management.

Legal content sits in what Google calls YMYL — Your Money or Your Life. That classification means Google holds your content to a higher standard than almost any other industry. If an AI writes a sloppy blog post about the best running shoes and gets a detail wrong, nobody gets hurt. If an AI writes a sloppy blog post about DUI penalties in Arizona and gets the sentencing ranges wrong, someone might make a real decision based on bad information.

Google knows this. Their quality raters are trained to scrutinize legal content for expertise signals that generic AI output simply doesn’t have — jurisdiction awareness, procedural accuracy, the kind of nuance that only comes from actually practicing in an area.

So when you feed ChatGPT or Claude a lazy prompt and publish whatever comes back, you’re not just getting mediocre content. You’re publishing something that Google’s systems are specifically designed to identify and suppress.

Five Ways Firms Sabotage Their Own AI Output

We see the same mistakes across almost every firm we audit. Not because the people are careless, but because nobody taught them how these tools actually work.

The context-free question. “Write a blog post about car accidents.” This gives the model no jurisdiction, no audience, no angle. You’re asking it to guess. It will guess wrong, or more accurately, it will guess generically — which for legal content is the same thing.

The kitchen-sink prompt. “Write a blog post about truck accidents, also draft three meta descriptions for my practice area pages, and suggest some email subject lines for our newsletter.” Three unrelated tasks, three different contexts, one mediocre attempt at all of them. The model’s attention gets split, and nothing comes out sharp.

The one-and-done. Firm sends a single prompt. Gets back a C-minus draft. Concludes AI doesn’t work for legal content. What they should have done: treat that first response as a rough draft and spend two more prompts refining it. The value is in the iteration, not the first pass. It’s a conversation with a research assistant, not a slot machine.

Telling it what you want but not what you don’t want. “Do not include specific legal advice. Do not assume jurisdiction. Do not use a casual tone. Do not reference case outcomes without citation.” These negative constraints are just as important as the positive instructions, and most people skip them entirely.

Not giving it permission to say “I don’t know.” This is the one that caused our Texas comparative negligence mistake. AI tools will confidently generate plausible-sounding legal information even when they’re uncertain. One line fixes most of this: “If you’re not sure about a specific statute, case, or legal standard, flag it for verification rather than guessing.” We put this in every single prompt now.

The Tactics That Actually Changed Our Output Quality

We tried a lot of prompt engineering advice. Some of it is academic fluff. These seven things made a measurable difference in the content we produce for law firm clients.

Give Context Before You Give the Task

Think of it this way. If you hired a contract attorney and on their first day said “write something about negligence,” they’d stare at you. If you said “We’re a PI firm in Houston. We need a 1,200-word blog post explaining comparative fault in Texas for people who were just in car accidents and are googling their options,” they’d get to work.

AI tools are the same. Before you tell them what to write, tell them who they’re writing for, what practice area, what jurisdiction, and what the reader’s situation is. We call it the 3 W’s — what’s the task, who’s the audience, what format do you need — and it goes at the top of every prompt.

The difference in output quality is stark. Without context, the model pulls from everything it knows about negligence across all 50 states and produces something blandly correct but specifically useless. With context, it narrows to the intersection that actually matters for your firm.

Be Relentlessly Specific

This is the single change that moves the needle the most.

What doesn’t work: “Write about estate planning.”

What does: “Write a 1,000-word blog post for a Texas estate planning attorney’s website. Target audience is married couples, 45-65, with kids and a net worth between $1M and $5M. Topic: why a will alone isn’t enough and when a revocable living trust makes sense. Include Texas-specific probate context. Tone should be authoritative but accessible — no legal jargon without an explanation. Don’t include specific tax figures that might be outdated by the time we publish.”

The first prompt gets you a Wikipedia summary. The second gets you something an attorney can review and refine in 20 minutes.

Break Complex Tasks into Steps

We learned this the hard way on a 3,000-word guide about filing personal injury claims in California. First attempt: one big prompt, one big mess. Timelines were wrong, the comparative negligence section contradicted the damages section, and the FAQ answers were recycled from the body content.

What works: break it into four or five sequential prompts.

1. “Outline the key stages of a personal injury claim in California, from incident to resolution.”
2. “For each stage, explain what the claimant should expect and what deadlines apply under California law.”
3. “Add a section on comparative negligence under California Civil Code Section 1714 and how it affects compensation.”
4. “Write a practical FAQ section based on the questions a first-time personal injury claimant would actually ask.”

You verify each stage before moving to the next. The model builds on confirmed-accurate content instead of compounding errors across 3,000 words. It takes longer. The output is incomparably better.

Tell It Exactly What Format You Need

Left to its own devices, the model will give you whatever structure feels most probable. For law firm content, that usually means a wall of text with inconsistent heading levels. Specify exactly what you want.

Lazy version: “Give me FAQ content about DUI charges.”

Useful version: “Generate 8 FAQ questions and answers about DUI charges in Florida. Format each as a question in an H3 heading followed by a 75-100 word answer. Write at an 8th-grade reading level. Each answer should reference a specific Florida statute or procedural rule where relevant. Frame answers as general information, not legal advice.”

That output goes straight into your schema markup with minimal editing. You’ve eliminated an entire revision cycle by being specific upfront about structure.

Make It Show Its Work

This is the technique that made the biggest difference for legal accuracy. Researchers call it chain-of-thought prompting. In practice, it means asking the model to reason through a problem before giving you the answer.

Without: “What are the grounds for contesting a will in New York?”

With: “Walk me through each legal ground for contesting a will in New York under EPTL Article 3. For each ground, explain the legal standard, what evidence is typically required, and what the practical threshold for success looks like. Think through this step by step.”

The second version shows you the model’s reasoning. When it gets something wrong — and it will, eventually — you can see exactly where the logic broke down. You’re not just checking the final answer; you’re auditing the thought process. For YMYL legal content, that visibility is the difference between catching an error before publication and publishing it.

A 2022 study by Kojima et al. found that even appending “let’s think step by step” to a prompt measurably improved accuracy on reasoning tasks. For legal content, where precision actually matters, we consider this non-optional.

Show It Examples of What Good Looks Like

This is called few-shot prompting, and it’s the fastest way to get consistent tone and formatting across your content.

Say you need meta descriptions for 15 practice area pages. Instead of describing what you want, show the model two examples from your own site:

• “Injured in a truck accident in Houston? Our attorneys have recovered $50M+ for Texas accident victims. Free consultation — no fee unless we win.”
• “Facing criminal charges in Harris County? Former prosecutor now defending clients. 20+ years of experience. Call for a free case review.”

Then: “Write meta descriptions in this same style for these pages: [list]. Keep each under 155 characters. Include a call to action. Don’t include phone numbers. Reference the specific practice area and location.”

The model mirrors your examples. The output is on-brand from the first draft, and you spend your editing time on substance instead of reformatting.

Layer Everything Together on Real Tasks

None of these techniques work in isolation the way they work together. Here’s a real prompt structure we use for client content:

Context: “You’re writing for a criminal defense firm in Phoenix, Arizona. They handle DUI, drug crimes, and assault. The reader just got arrested and is searching for what happens next.”

Specific task: “Write a 1,200-word blog post about what happens after a DUI arrest in Arizona — the timeline from arrest through arraignment. Cover the Admin Per Se hearing, the 15-day deadline to request a hearing with the MVD, and the difference between a misdemeanor and aggravated DUI under ARS 28-1383.”

Format: “H2 headings for each stage. Practical FAQ at the end, 4-5 questions. 9th-grade reading level.”

Constraints: “No specific legal advice. No guaranteed outcomes. Reference Arizona statutes where relevant. If you’re uncertain about a procedural detail, flag it for attorney review instead of guessing.”

Reasoning: “Think through the DUI process chronologically before you start writing.”

Pull any one of those layers out and the quality drops. We’ve tested it. The context layer alone cuts generic filler by about half. Adding constraints on top of that eliminates most of the compliance issues. The reasoning instruction catches the factual errors that slip through everything else.

Build a Prompt Library So You’re Not Starting from Scratch

The firms that get real value from AI don’t reinvent the wheel every Tuesday. They build a shared library of prompt templates for the tasks they repeat:

Blog post template — practice area, jurisdiction, audience, word count, formatting, voice notes, and the “flag uncertainty” instruction baked in.

Meta description template — character limit, CTA style, geographic targeting, and 2-3 examples of approved descriptions for tone matching.

FAQ template — schema-ready format, answer length targets, reading level, and accuracy constraints.

Email sequence template — audience segment, funnel stage, desired action, and bar advertising compliance guardrails.

We keep ours in a shared doc. When a new team member starts producing content for a law firm client, they open the template, fill in the practice-specific details, and get output that’s already 80% of the way to publishable. Without the template, the same person produces output that needs twice as much editing.

Attorney Review Isn’t Optional — It’s the Whole Point

We’ll be direct about this because we’ve seen firms try to skip it: no matter how refined your prompts are, an attorney has to review every piece of legal content before it goes live.

Not because AI is bad at writing. Because AI doesn’t have a bar card, doesn’t carry malpractice insurance, and doesn’t know that the judge in your county handles preliminary hearings differently than the statute suggests. Google’s quality systems are looking for exactly the kind of expertise signals that only come from someone who actually practices law — specific case references, jurisdiction-aware procedural knowledge, the practical “here’s what this actually looks like” detail that no training dataset contains.

ABA Formal Opinion 512 laid out the framework. State bars are following. The ethical obligation isn’t “don’t use AI.” It’s “don’t publish AI output without competent human oversight.” Your prompt library makes the AI’s drafts better. Attorney review makes them publishable.

For more on where the lines are, see our guide on how law firms can use generative AI without getting penalized.

Where This Leaves Your Content Operation

We produce more content now than we did before we integrated AI tools. It’s also better content — not because the AI is smarter than our writers, but because the AI handles the structural and research grunt work and our writers spend their time on the parts that actually require a human brain: voice, accuracy, expertise, the stuff Google’s systems are designed to reward.

The bottleneck was never the technology. It was always the instructions. A good prompt gets you a draft worth editing. A bad prompt gets you a draft worth deleting. The seven tactics above are the difference, and they compound — once your team internalizes them, every piece of content gets faster and better to produce.

Pick one tactic. Use it on your next prompt. See what changes. Then stack another on top. That’s the whole process. Our content strategy guide covers how this fits into a broader publishing operation.