Part I – Prompts with Purpose: Unlocking Advanced AI Tactics for the Criminal Defense Advocate

How Deep Prompt Engineering Can Reshape Trial Strategy, Narrative Framing, and Sentencing Advocacy

Sep 05, 2025

The legal field runs on language, statutes, cross-examinations, narratives, and closing arguments. Now, with generative AI, we have a tool that runs on language too. But to get precision, coherence, and strategy from AI, lawyers must learn to speak its dialect. That dialect is prompting.

Many criminal defense attorneys interact with AI through casual queries, without sophistication or forethought. Advanced prompting isn’t just about getting better answers. It’s about constructing better questions, more intentional workflows, and persuasive output that mirrors the rhythms of real advocacy.

This first installment in a three-part series explains why prompt engineering matters for criminal defense, revisits foundational techniques (with a summary chart from prior articles), and introduces a new technique: Step-Back Prompting - a strategy that breaks from the instinct to "ask and receive" and instead encourages staged legal reasoning before execution.

Revisiting Core Techniques from Earlier Work

In my four NACDL Champion Magazine articles (see links below) I covered many basic and more advanced prompting techniques that you can use to facilitate your litigation practice.

These articles explored essential techniques such as role prompting, contextual grounding, reasoning structures (including Chain of Thought), zero- and few-shot prompting, task segmentation, persona prompting, question refinement, and alternate approach generation.

Follow this link to download a PDF Chart of previously described prompts for criminal defense. The chart is adapted from those materials, summarizes the core techniques already covered.

The techniques introduced below are meant to build upon and embellish this foundation, offering novel methods for refining persuasion, aligning AI reasoning more closely with the patterns of advocacy, and producing more useful content overall.

Step-Back Prompting: The Wisdom of Withholding

Sometimes the most persuasive question is the one you don’t ask—yet. Step-back prompting encourages the model to pause before it performs. It prompts first for principles, then for application.

Step-back prompting works by encouraging a staged inquiry. Instead of diving straight into the final product—like a closing argument or full legal analysis—you begin by prompting the model to surface the foundational principles, heuristics, or case-specific variables that should inform that final output.

This approach is especially useful when a task involves scientific evidence, cross-disciplinary issues, or ambiguous legal standards. By generating the framework first, you can guide the model more precisely and avoid premature or unfocused answers.

You don’t begin with the opening statement—you begin with:

“What are the core factors that undermine or reinforce the credibility of a forensic blood test in DUI cases involving suspected contamination or delayed analysis?”

You would then use the output generated by this prompt to continue your inquiry by identifying which of the reliability factors are most relevant to your case, and then incorporating those into follow-up prompts.

For example, if the model highlights sample degradation or chain-of-custody concerns, you can direct the model to formulate specific cross-examination questions, draft motions in limine, or generate alternative narrative framings based on those concerns. This staged use of prompts ensures that each subsequent output is grounded in a coherent evidentiary and argumentative foundation.

Examples in Criminal Defense:

Challenging the admissibility of a blood test in a vehicular manslaughter case
First prompt:

“What are the common factors that affect the reliability of forensic blood alcohol results, particularly in cases involving potential sample degradation or improper storage?”

Then follow with:

“Based on those factors, generate five cross-examination questions that could raise doubt about the reliability of the blood test without appearing overly aggressive toward the analyst.”

Constructing a defense theory around mistaken identity in an armed robbery case
First prompt:

“What are the psychological and procedural factors that can lead to eyewitness misidentification in high-stress, short-duration criminal events?”

Then follow with:

“Draft voir dire questions designed to uncover potential juror overreliance on eyewitness testimony in the absence of corroborating forensic evidence.”

Developing a motion to suppress a warrantless cell phone search in a drug possession case
First prompt:

“What are the key constitutional arguments under the Fourth Amendment regarding warrantless searches of cell phone data post-Riley v. California?”

Then follow with:

“Draft a motion to suppress that integrates the reasoning from Riley and applies it to a scenario where officers accessed call logs and GPS data without a warrant.”

These layered prompts allow the practitioner to build their case in stages—developing the conceptual groundwork before generating advocacy strategies.

You’re not just simulating a script. You’re architecting intention.

Coming Up in Part II

In the next article, we’ll explore how to shape AI output more precisely by defining internal constraints and rhetorical purpose. You’ll learn how System Prompting, Automatic Prompt Engineering, and Self-Reflective Prompts can elevate your advocacy with greater control, tone alignment, and strategic revision.

Learn more about Barone’s AI-powered Criminal Defense Practice at www.baronedefensefirm.com, where our mission is to provide cutting-edge, compassionate, and relentless defense by combining the best of human advocacy with the smartest technology available.

AI on Trial

Discussion about this post