Case study

Using AI to Develop True Crime

Matt O’Brien, November 2026

What if a production company could take a true-crime story and hand an editor an assembled cut, rather than a folder of research and a blank timeline? That is the development problem I wanted to work on. The editor can improve the storytelling instead of starting from scratch.

I built a completely automated pipeline to turn source material into a development cut. The aim was to do it with a cheap model and a finely tuned set of agents, rather than make the newest model do all the work. I see this as the future of unscripted development.

Can we get to the first cut sooner?

There are two ways in. A news article supplies the reported story. An existing episode and its timed subtitles also show how that story was told. Both paths turn the source into decisions an editor can work with.

Start with a story

In the article workflow, I paste a news URL. The system extracts the article text and uses it as the basis for a scene-by-scene script. It exports structured JSON for Veo video generation, so the story breakdown becomes an input to production rather than another document to retype.

A news article is a starting point, not a script. The article path extracts the reported story, breaks it into scenes and exports structured JSON for video generation. The source still needs editorial review before its claims become a programme.

From 44 minutes to ten short episodes

The adaptation question is not how to chop up subtitles. It is how to take a self-contained, 44-minute true-crime story and reshape it into ten episodes of about three minutes each.

A schematic 44-minute broadcast-hour shape. Handwritten timings are development prompts to test, not measured averages or research findings.

We study the source episode’s cadence: where it turns, what each reveal changes, and what it withholds. Then we redistribute those beats across the shorter run. Timed pictures and transcripts keep the analysis anchored to the source; they support the adaptation rather than dictate its cuts.

On camera expert shot: still from True Crime Microdrama Ep. 1

On Camera Expert

Interview shot: silhouetted interviewee, still from True Crime Microdrama Ep. 1

Interview

Dialogue shot: still from True Crime Microdrama Ep. 1

Dialogue

Sample shot types: On Camera Expert / Interview / Dialogue. Stills from True Crime Microdrama Ep. 1, the episode built with the pipeline, show how it classifies a face on screen. The scenes are dramatized; these are not the real people in the source story.

Each short episode needs its own opening hook, a meaningful turn and an unresolved finish that earns the next episode. This is the adaptation approach, not a claim that all ten episodes have already been produced.

Four distinct development moments: the opening hook, a change in understanding, the reveal and the turn. Original schematic illustrations; no real case events depicted.

Torn development note: the unresolved question that carries a microdrama into its next episode

Microdrama cliffhanger: resolve one beat, leave the next consequence unresolved. The question on this torn note is fictional, not a reported case event.

Genre experience inside the pipeline

I honed the agents with true-crime genre experience. They make agentic decisions about speaker versus narrator, B-roll versus interview, which story point belongs next, and where an episode should leave us wanting more. The output is an assembled development cut an editor can upgrade.

The point of the test was complete automation with a cheap, lower-tier model. The early proof used Veo 3. I do not consider that a frontier model now, and it is not the ceiling. It shows how much of the work can come from the pipeline rather than the model alone.

For true crime, human review also has to cover factual accuracy, the treatment of allegations and the distinction between reporting and generated reconstruction. An automated scene plan is not evidence. A development cut is not a cleared programme.

The economics are about a slate

The business question is what happens when a self-financed production company uses a cheaper model to make around 100 short episodes and distributes them on social. At that scale, the numbers can start to work. That is the opportunity I want to test, not a claim that this prototype has already proved a return.

A finely tuned pipeline makes the model one production choice within the process. The documented design uses cheaper generation for B-roll and narrated sequences, and reserves more expensive generation for critical scenes. Historical prototype prices are not a current quote.

Generation cost is only part of the calculation. Editorial time, fact-checking, rights, revisions and distribution still count. The useful comparison is the cost of a repeatable, reviewed episode, not the price of a single impressive clip.

The question is whether a small company can repeat that process across a slate: sourced material, agent decisions and an assembled cut, with human editorial review before release.

What I built with the pipeline

This is the true-crime micro-drama I built using the pipeline. It is a little different from a straight condensation of an existing episode, but it uses the same structure and some of the same decisions.

Full official Vimeo newspaper-cover composition

True Crime Microdrama Ep. 1 (AI-Generated). 2:55. Public sample credits include Veo 3.1.