Connectors, skills, and projects. Prosecutor use cases.
The doors from Part 2, in practice: how a model reaches your email and your research tools, how a saved procedure turns one lawyer’s process into the office’s, and what a case-screening skill does with a real file.
Plug the model into your systems.
Connectors plug your AI model into outside data: your email, SharePoint, research databases, and more. Instead of copying and pasting, the AI goes and gets the information itself. Click a system to see the kind of question you can ask.
The AI reads, and sometimes acts, straight from the source system.
One question can reach email, files, calendar, and a research database at once.
The AI inherits your access. Whatever the connector reaches, the AI provider may process. Check before you connect.
Under the hood: most connectors use MCP (Model Context Protocol), an open standard. Build one and any AI tool that supports it can use it.
The office manual, not the conversation.
A skill is a saved set of instructions that teaches the AI to do one task your way, every time.
Explaining a task to a new ADA out loud.
Useful once, gone tomorrow.
Writing the procedure down once, with the statutes and forms attached.
Anyone in the office can follow it, the same way, every time.
Instructions
When to use it, the steps to follow, and what a good answer looks like.
Reference files
The sources it should rely on: statutes, pattern jury instructions, office policy.
Templates
The finished format: a charging memo, a table, a Word document ready to review.
Three skills in Claude. The prosecutor analysis calls the other two.
Each skill is a set of instructions plus the reference files it reads. All three ship in one plugin and install together. The analysis finds both libraries at the start of every run, and a companion defense-analysis skill reuses the same two libraries.
Screens each charge for probable cause. Reads the statute text and loads the element language.
Chapter 14 sections
Instructions, linked to the statutes by index
Never state elements from memory. Load the N.C.P.I. instruction and use its words.The skill’s first rule. Statute text is read from G.S. Chapter 14 for every charge; anything outside the library is flagged, not guessed.
Why this rule exists. Six prosecutors who reviewed AI summaries of ten real cases most often flagged paraphrased element language. Their criticisms became the design rules.
Source: Cuffney, Ethically Implementing Generative AI in Prosecution (Northwestern University, 2025); ncleg.gov; NC School of Government.
Five case files. One prompt. Same skill every time.
In the talk I walk through five AI-drafted prosecutorial summaries built from redacted public-records case files. The case details stay in the room. Here is how the demo was run, and what held up across every file.
A redacted case file
Incident report, case summary, and every officer narrative and supplement. Scanned PDFs, 5 to 34 pages, with degraded OCR and black redaction bars.
The skill, and one line
Run a prosecutorial summary on the attached case file. Names are redacted. Use placeholders.Skill loads: N.C.P.I.—Crim. · G.S. Ch. 14 · issue checklists · summary template
Nothing about the charges or the law is in the prompt. The skill goes and gets it.
A six-section draft
A Word document, labeled a draft. Elements quoted verbatim, every fact traced to a report quote, defenses always analyzed, and a verification pass before delivery. The charging decision stays with the prosecutor.
The prompt is identical across cases. Differences in output are driven by the file, not by hand-tuning.
What the model did well, every time.
The strengths are consistent across files because they are encoded in the skill, which was itself written from the prosecutors’ criticisms in the thesis study.
Led with the case-killer
Every summary opened with one or two case-critical flags before a single element table.
Quoted, never paraphrased
Elements came verbatim from the N.C.P.I., with revision years. Proof came as report quotes with narrative numbers. Reconstructed OCR was tagged, not repaired.
Found the discrepancies
Witness descriptions that don’t match, force described three different ways, a hedged “believed” next to a “positively identified.” The thesis’s biggest miss became a strength.
Always analyzed defenses
Self-defense, accident, alibi, mere presence. Addressed in every file, including the ones where the answer was “not supported by these facts.”
Offered the fix
Not just the flaw: an alternative charging theory, a lineup under the Eyewitness Identification Reform Act, a certified prior judgment, lab work that was never requested.
Stayed in its lane
Every document is labeled a draft, cites RPC 3.8 and ABA Standard 3-4.3, runs a stated verification pass, and explains its own length when it ran over.