It is 11:40 p.m. on a Thursday. You have just cleared a domestic disturbance call, your third of the shift. The interaction lasted eighteen minutes, your body-worn camera (BWC, the camera clipped to your chest that records video and audio throughout the contact) captured all of it, and you have forty minutes before end of watch to finish the incident report. You open the AI-assisted drafting tool your agency deployed six weeks ago. You type: "Write my report." The model obliges. It produces three paragraphs of clean, confident, professional prose that sounds exactly like a police report should sound. You read it quickly. It seems right. You submit it.

That prompt just cost you more than you know.

Not because the tool is bad. The tool may be excellent. Axon's Draft One and similar platforms have shown testing officers an 82% decrease in report-writing time, and that number represents real hours returned to patrol and community contact. The tool is not the problem. The problem is the three words you gave it to work with. "Write my report" is the weakest possible instruction you can issue to a generative AI model. It omits context, sets no accuracy standard, imposes no grounding constraint, and invites the model to fill every gap with a statistically plausible detail rather than a verified fact. This lesson is about closing that gap: how to prompt so the output is useful, accurate, and defensible when a defense attorney, a prosecutor, or a judge reads it.

Why the Prompt Is the First Line of Defense

A generative AI language model does not have a source of truth. It has a very large statistical memory of language patterns. When you give it a task, it produces text that is consistent with those patterns and with whatever input you provide. If your input is sparse, the model fills the gaps with what typically happens in incidents like this one. That is not verification. That is gap-fill, the first and most dangerous hallucination type in public-safety work.

Think of it this way. Imagine briefing a new officer who was not on scene: "Write up what happened at that domestic call." That officer has no footage, no CAD entry, no notes. They have only what they know about how domestic calls usually go. What they produce might be plausible, but it is not your account of this incident. It could be missing the specific statement the subject made that becomes the central disputed fact at trial. It could omit the detail about the broken lamp that establishes the physical contact. It could soften a resistance fact because domestic calls are often resolved without resistance and the officer is writing what usually happens rather than what happened.

That is what your AI does when you give it an empty prompt in a high-stakes context. The model is not lying. It is doing exactly what it was trained to do: produce fluent, contextually appropriate text. But fluent and contextually appropriate is not the same as accurate and grounded, and in a sworn report, accuracy and grounding are the only things that matter.

The prompt is not a convenience setting. It is the instruction that determines whether the AI drafts from your incident or from its training data's average incident.

The Three Habits That Separate Usable from Unusable

Every effective public-safety AI prompt rests on three habits. Learn these three and you will spend the rest of this lesson and the ones that follow building them into muscle memory.

Habit 1: Provide context. The model cannot use what you do not give it. Context means the incident data: the call type from your CAD (computer-aided dispatch, the system that receives and routes calls for service and maintains the incident log), the address, the time, the parties involved, the call notes that came across when you were dispatched, and most importantly, the source material you want the model to draw from. "Based on the following BWC transcript and CAD notes" is context. "Write my report" is not.

Habit 2: Set an accuracy standard. Tell the model explicitly that it must not add, embellish, or speculate. The phrase you will use again and again in this program is "only what the footage shows." That phrase is not decorative. It is a constraint that changes the model's behavior. Without it, the model treats your input as a starting point and fills from training patterns. With it, the model treats your input as a boundary it must not cross.

Habit 3: Force the model to the record. Your record in a public-safety context is the BWC footage, the CAD entry, and the case file in your RMS (records management system, the database where case reports, arrest records, and related data are stored and retrieved). Name the sources explicitly in your prompt. "Based only on the attached transcript of BWC footage, incident number 2026-04471, and the attached CAD notes" is grounding. It tells the model where the facts must come from. If a fact is not in those sources, the model should say so, not invent it.

Anatomy of a Strong Public-Safety Prompt

Let us build the prompt that should have opened that 11:40 p.m. drafting session, piece by piece.

The first element is the role instruction. You are telling the model how to behave: "You are a sworn law-enforcement report writer. Your function is to draft a factual incident narrative from the materials I provide." This is not theater. Role instructions constrain the model's output register to something that reads like a professional sworn account rather than a chatbot response.

The second element is the source instruction. "Draft this report using only the following materials: (1) the transcript of my BWC footage attached below, and (2) the CAD notes for incident 2026-04471 also attached below. Do not add information not present in these materials." This is the grounding constraint. It tells the model what the sources are and prohibits gap-fill.

The third element is the accuracy rule. "If any detail is unclear in the footage or absent from the CAD notes, note the gap explicitly with the phrase 'not captured on available footage' rather than inferring or estimating." This is the honesty instruction. Instead of letting the model paper over a gap with a plausible detail, you are asking it to flag the gap so you can fill it yourself from your own recollection, noted as such in the final report.

The fourth element is the format instruction. "Format the narrative in past tense, third person for subjects and first person for officer actions, matching agency report style. Include time markers from the footage timestamp." This produces output that fits directly into your RMS rather than requiring extensive reformatting.

Put together, the full prompt looks like this:

"You are a sworn law-enforcement report writer. Draft a factual incident narrative for a domestic disturbance call using only the following materials: (1) the BWC footage transcript attached below, and (2) the CAD notes for incident 2026-04471. Do not add any information not present in these materials. If a detail is unclear or absent, write 'not captured on available footage' rather than inferring. Format in past tense with time markers from footage timestamps."

That prompt will produce a fundamentally different result than "Write my report." It will produce a draft that is grounded, that flags its own gaps, and that you can verify against the footage in a fraction of the time you would need to verify a free-form draft.

Context Is Not Optional, It Is the Material

One of the most common mistakes new users of AI drafting tools make is treating the transcript or CAD notes as optional attachments: "I'll just describe the incident and let the model fill in the details." This approach misunderstands what the model is doing. When you describe an incident conversationally, the model combines your description with its training patterns to produce a narrative. The more detail you omit, the more the model fills in from training. Gaps become invented details. Omitted facts become plausible substitutes. The report reads well, but it is increasingly a blend of your incident and the model's statistical memory of similar incidents.

The fix is not to write a longer conversational description. The fix is to give the model the actual source material. Most modern AI-assisted drafting tools for law enforcement accept BWC transcripts, CAD exports, and case file attachments directly. Use them. When the model has the transcript in front of it, it is not filling from training memory. It is organizing and synthesizing text that came from your actual incident. The gap-fill problem does not disappear, but it shrinks dramatically.

If your tool does not accept attachments and you are working with a general-purpose AI assistant, paste the relevant text directly into the prompt. It is more typing, but it is the difference between a report grounded in your incident and a report grounded in the model's guess about your incident.

The "Only What the Footage Shows" Constraint

This phrase is the most important single tool in your prompting kit. It is worth understanding why it works so you can apply it correctly in situations the template does not cover.

A language model generates text by predicting what comes next, given everything that came before. Without constraints, the model's "what comes next" is guided by the full distribution of its training data: every police report, news article, legal document, and narrative it has ever processed. That distribution is informative, but it is not your incident. When you include "only what the footage shows" as an explicit instruction, you shift the model's reference point. You are saying: the footage transcript I have attached is the authoritative text. Draw only from it.

The constraint does not make the model infallible. Models can still make mistakes even with a grounding constraint. They can misread a timestamp, mishear a word in a transcript, or misconstrue an ambiguous phrase. But "only what the footage shows" eliminates the largest class of errors: the invented detail that was never in the footage at all. It tells the model that speculation and inference are outside the scope of this task.

You should apply this constraint, or a functional equivalent, in every prompt where you are asking the model to produce text that will appear in a sworn report, a CAD entry, a supplemental report, or any document that will be disclosed to a prosecutor or defense attorney. This includes:

• Initial incident report narratives drafted from BWC footage.
• Supplemental reports summarizing follow-up investigation.
• Evidence log descriptions produced from footage or photographs.
• CAD entry summaries drafted from dispatch audio transcripts.
• Interview summaries produced from recorded interview transcripts.

It does not apply to administrative or internal tasks where accuracy to a specific source document is not required: drafting a shift briefing, summarizing a policy document, or generating a training scenario. In those contexts, the model's general capabilities are fine. In evidentiary contexts, the constraint is mandatory.

What the Model Should Say When It Does Not Know

A well-prompted model will sometimes tell you it cannot find a specific detail in the provided materials. That is a feature, not a failure. When the model writes "not captured on available footage" or "not present in provided CAD notes," it is doing exactly what your prompt instructed it to do: flagging a gap rather than filling it. That flagged gap is your cue to do one of three things.

First, check the footage yourself. Sometimes the detail is there and the model missed it, misheard it, or was looking at the wrong timestamp window. Go back to the source.

Second, add the detail from your own recollection, clearly attributed. If you recall a fact that was not captured on footage, you can include it in the report, clearly identified as based on officer recollection: "Observed by the responding officer, not captured on BWC due to camera angle." That attribution is honest, protective, and legally appropriate. What you cannot do is let the model invent the detail for you.

Third, leave the gap if the detail is not necessary. Not every incident report requires every conceivable detail. If the model cannot find it and you cannot verify it, the report may be complete without it.

The wrong response to a gap flag is to tell the model: "Just write something plausible." That instruction undoes the grounding constraint entirely and returns you to the gap-fill problem. Never paper over a flagged gap with an AI-generated invention.

Grounding the Model to CAD and RMS

The CAD (computer-aided dispatch) entry is a second source of truth distinct from the BWC footage and often more important for establishing the official record of the call. The CAD entry includes the call type, the received time, the dispatch time, the arrival time, the address, and the notes entered by the telecommunicator (the dispatcher who answered the call and coordinated the response). These are objective timestamps and classifications. They do not depend on the quality of the BWC audio or the clarity of the footage. They are the official record of when and how the call was handled.

When you are prompting for a report narrative, include the CAD export alongside the BWC transcript. This is not redundant. The CAD entry provides the structural spine of the report: times, addresses, and call type. The BWC footage provides the narrative content: what was said, what was observed, and what actions were taken. Together they produce a draft that is both factually anchored and narratively complete.

The RMS (records management system) is where the finished report lives. Most RMS platforms have specific field requirements: a narrative field, an offense code, a suspect description field, a property loss field, and so on. If your agency's AI tool produces structured output, you can include RMS field labels in your prompt to get output that maps directly to those fields: "Please format your output with separate sections for: Narrative, Subjects, Property, and Evidence, matching the field structure of our RMS."

Grounding to CAD and RMS is not a technical nicety. It is the difference between a draft you can verify in twelve minutes and a draft that requires forty minutes of reformatting before you can even begin verification. The officer who invests thirty seconds in a well-structured prompt buys back time at every downstream step.

The Prompt Template You Should Memorize

This program will give you many prompt templates tailored to specific use cases. But the foundational template for any report-writing task is worth internalizing so thoroughly that you can reconstruct it even when you are tired, rushed, or working from memory at the end of a twelve-hour shift.

The template has five slots:

[Role] You are a sworn law-enforcement report writer.
[Source] Draft this report using only: (1) [source 1] and (2) [source 2].
[Accuracy rule] Do not add information not present in these materials. Flag gaps with "not captured on available footage."
[Format] Format in [tense/person/style], include time markers from footage.
[Material] [Paste transcript / attach file here].

Five slots. Two to three minutes to fill. Forty to sixty minutes of downstream verification saved because you started from a grounded draft rather than a free-form AI narrative.

Practice filling those five slots for a real incident before the end of this lesson. You do not need a live AI tool to practice. You can write the prompt on paper, identifying what goes in each slot for a recent call you handled. The exercise of identifying your sources and your accuracy constraint before you even open the tool is itself training for the habit that will protect your reports.

Advanced Prompt Moves for Public Safety

Once the five-slot template is habitual, there are several advanced moves that produce better results in specific public-safety contexts.

The tense and attribution instruction. By default, models sometimes mix tenses and switch between first and third person mid-narrative. Adding an explicit instruction ("Use past tense throughout. Refer to the officer in first person and to all other parties by role label: Subject 1, Witness 1, etc.") prevents the kind of inconsistent prose that draws scrutiny during review.

The legal standard instruction. For incidents likely to result in arrest reports, criminal complaints, or testimony, add: "Draft the narrative to support the specific offense elements of [charge], as defined in [jurisdiction code section]." This focuses the model on the legally relevant facts rather than producing a general description that may omit an element required for the charge. This is not legal advice; it is a focusing instruction that keeps the narrative on point.

The completeness check instruction. After producing the initial draft, issue a follow-up prompt: "Review the draft you just produced. For each factual claim, identify whether it is directly supported by the provided footage transcript or CAD notes, or whether it represents an inference or addition. List any inferences or additions separately." This meta-review instruction asks the model to audit its own output against the grounding constraint. It is not a substitute for your own verification pass, but it surfaces the model's own gaps and inferences before you begin your review, which makes your verification faster and more targeted.

The Brady and Giglio flag instruction. Brady v. Maryland requires the prosecution to disclose exculpatory evidence. Giglio v. United States extends that obligation to evidence that could impeach a witness, including the testifying officer. For incidents where you anticipate contested testimony, add to your prompt: "Flag any fact in the provided materials that could be considered exculpatory or that represents an inconsistency between the footage and the initial dispatch notes." This does not replace your own Brady and Giglio review, but it gives you a starting list of items to scrutinize in your verification pass.

These advanced moves require you to know your use case before you prompt. That knowledge comes from field experience. The more you understand what makes a report strong in court, the better your prompts will be. Prompting well is, in the end, an extension of knowing how to write a strong report. You cannot prompt for something you do not understand.

What Good Prompting Cannot Fix

Prompting is the first line of defense, not the only one. Understanding the limits of good prompting is as important as understanding its power.

No prompt will fix a BWC transcript that is inaudible. If your footage has a critical ninety-second segment where the audio is obscured by wind noise, sirens, or overlapping voices, even a perfectly grounded prompt will produce a draft with a ninety-second gap. That gap will show as "not captured on available footage," which is honest. But the underlying problem, the gap in the record, remains. Prompting cannot manufacture evidence that does not exist.

No prompt will catch every hallucination. Models can make mistakes even with a grounding constraint in place. They can misread timestamps, misconstrue ambiguous phrases, or confuse two subjects whose descriptions are similar in the transcript. The grounding constraint reduces hallucinations dramatically, but your verification pass remains mandatory. The lesson on verification, coming up next in this chapter, will walk you through the specific checks that catch what the best prompt still misses.

No prompt eliminates the authorship obligation. When you submit a report that you generated with AI assistance, you are the author. The CJIS (Criminal Justice Information Services) Security Policy, maintained by the FBI, governs how criminal justice information is collected, processed, and stored. Under CJIS and under your agency's policy, the sworn narrative in your report is your sworn account. "The AI wrote this" is not an answer to the deposition question "Officer, did you write this?" You wrote it. You reviewed it. You adopted it. The AI assisted. That accountability does not change with a better prompt, and it should not.

What good prompting does is give you a better starting draft, a more targeted verification task, and fewer invented details to catch and correct. It makes the hours shorter and the risk lower. It does not make the risk zero and it does not shift accountability away from the sworn officer who signs the report.

Key Takeaways

• The weakest AI prompt in law enforcement is "write my report." It produces a fluent, confident draft grounded in training patterns rather than your incident, and every gap becomes an invented detail.
• Every effective public-safety AI prompt rests on three habits: provide context (source material), set an accuracy standard ("only what the footage shows"), and force the model to the record (BWC transcript, CAD entry, and RMS case file).
• The five-slot template (role, source, accuracy rule, format, material) takes two to three minutes to fill and eliminates the largest class of report errors before they start.
• The phrase "only what the footage shows" is a constraint that shifts the model's reference point from its training data to your actual incident record. It should appear in every prompt that produces text for a sworn document.
• When the model flags a gap with "not captured on available footage," that is a feature. The correct response is to check the source, add from your own recollection with attribution, or leave the gap. Never tell the model to invent a plausible detail.
• Advanced prompt moves (tense and attribution instructions, legal standard instructions, completeness check prompts, Brady and Giglio flag prompts) extend the basic template to specific high-stakes use cases.
• Good prompting reduces hallucinations but does not eliminate them. The verification pass against the footage and the CAD entry remains mandatory regardless of how well you prompted.
• Authorship and accountability remain with the sworn officer regardless of how the draft was generated. A better prompt does not change who signs the report or who answers for its accuracy under oath.