Of all the things involved in turning a woman's life into a 60-second reel, I assumed that the writing would be the easy part. Surely telling a good story in 150 words is exactly what a large language model should be good at; yet it proved surprisingly difficult. Draft after draft suffered from common AI weaknesses: a tendency to use hyperbole and inspirational language, to generalize, follow generic founder arcs ("built in a basement"), and focus on morbid details. (For this effort, I was using Sonnet 4.6, which I found to be better — more grounded, more true to facts, less inventive — than Opus 4.7 at creating longer bios.) Getting to something publishable took a significant amount of work. That said, the human in this story also found writing good short story arcs surprisingly challenging, so the effort spent on getting the system to do it decently was well worth it. Here's some context, then what I did.
I've been publishing short biographies of notable women for a while now, on a website called Tycoona. Why notable women? Because in each field there are so many women who contributed so much and are little known. I've never understood why Corita Kent, the pop-artist nun, isn't as famous as Andy Warhol, why Hetty Green, the Gilded Age value investor called both the "Witch of Wall Street" and the "Queen of Wall Street," disappears into history behind Benjamin Graham and his protege Warren Buffett. The problem I faced is that no one was seeing my bios, so I decided to make short videos or reels in hopes of increasing my reach.
Creating the technical infrastructure for the reels on top of my existing system was fun and relatively straightforward. Each reel consists of a series of "beats," and Remotion turns those beats into a vertical reel, complete with on-screen text and royalty-free images & attributions pulled from Wikimedia, Flickr, or Library of Congress. For content, the beat generator relies on the knowledge base of validated facts that my system creates for each person: quotes, a timeline, achievements, relationships, legacy, additional facts, and "contested items." Here are the 3+ things I did to make the arc writer work (better).
Screenwriter Persona/Skill
First, I created a screenwriter persona, a global Claude skill. The Screenwriter specializes in writing compelling stories for short videos, anchored in facts, where the viewer always knows where they are in time. The Screenwriter knows that each reel needs an arc (a hook, a turn, a payoff), rather than a list of accomplishments strung end to end. A "context" file in my project tells the global screenwriter skill where the project's arc-generation prompts live, which database fields hold the validated facts it must draw from, and what other skills/personas it has available to help.
I had the screenwriter rewrite the beat creation prompts, which I then edited myself and exposed in the UI. In the process, my "chief architect" skill picked up that the beat creator was getting contradictory instructions from two different places. We rationalized these into a single, pre-existing voice.md file which contains key principles: let the facts speak for themselves, avoid hyperbole and generality, choose mechanics over hagiography, significance over morbidity. We also refined the factuality prompt: not only should each beat have validated facts, but the entire beat must be factual.
These changes definitely helped. I was able to get stories, rather than lists of facts. Consistent access to the voice.md file helped get rid of some of the model's morbid tendencies, for example, its insistence on mentioning that Marie Curie is buried in a lead box and her lab notebook must stay in a lead box for 1500 years, due to radioactivity. Finally, the "fact" prompts kept the model from inferring things from facts that were not documented.
Despite this progress, the story arcs lacked luster. The arc generator retained many of the "bad habits" of LLMs: a tendency to lean on generic narratives, use vague language, and omit compelling details.
Explicit Story Arc
While iterating over the various drafts, it became clear that not only does a single life have numerous possible arcs, but that surfacing the chosen arc would help the system generate a better story.
A good example is the scientist Rosalind Franklin, who, in popular narrative, was cheated out of a Nobel by Watson and Crick. What's more, the popular grievance narrative is not 100% true — Franklin's work and Watson and Crick's was funded by the same research organization, and the data sharing might have happened in the legitimate pursuit of beating the Americans to the punch. Quite a few different arcs are possible for this character:
"Scientist whose 'Photo 51' of DNA was shared with rivals without her consent."
"Woman behind four men who won two Nobel prizes."
"X-ray crystallography expert who, by uncovering the structure of molecules, made significant contributions to the fields of materials science, biology, and virology before the age of 37."
I am uninterested in writing grievance narratives and more interested in really understanding what these women did and how. So to start getting the stories I wanted, I made the story arc explicit in my system and let the beat generator work from a given arc. To save myself time, I made it possible for the system to generate five different arcs for me to choose from, though I generally write my own.
Add Facts to the Arc
Each life has unexpected and little-known details that make for a juicy story. Katherine Johnson, the brilliant NASA mathematician, had a 13-year career "gap" between leaving grad school and starting at NACA (later NASA) during which she raised 3 daughters. And her return to work appears to have been precipitated by her first husband's cancer diagnosis. Judy Faulkner, the founder of health records giant Epic Systems, was one of the first women to do graduate work in the newly created field of Computer Science, at UW-Madison, in 1965. And her famously still-private company's commandments begin with: "Do not go public. Do not acquire or be acquired. Software must work."
Any good story needs some details that a reader can hang onto, and the model is not great at picking the good ones. In any case, the trick to getting a more compelling arc out of my system was adding facts to the arc. With this scaffolding, the system is able to create concise, coherent, occasionally compelling story arcs for reels.
Having to Leave Some Things Out
For a human caught up in a story, a painful part is what to leave out. Stephanie Shirley is a little-known British computer programmer who, after experiencing sexual harassment and glass ceilings at work, started a software engineering firm in 1962 employing women caregivers who needed flexible hours and work from home. As the firm grew, she gave controlling interest to her employees, then sold the firm for close to half a billion dollars and gave her fortune to charity. As it happens, she also arrived in England from Germany in 1939 at the age of five on the Kindertransport with her nine-year-old sister. She expressed her lifelong motivation in a single line: "I wanted my life to have been worth saving." But all of these narratives will not fit into a 150-word arc. Some have to be left out, and that's sad. And some can't even make it into a 1000-word biography.
The Takeaway
LLMs can generate content quickly, but not necessarily well. Their performance depends on context — in this case, a knowledge base and specific instructions on how to use it, a set of style guidelines, an articulated story arc, and selected details that support that arc. Get that context right, and the LLM can help produce something worth publishing. Thanks to my reel generator — which admittedly relies heavily on human inputs but undoubtedly also saves me hundreds of tedious hours — more than a quarter million people across Instagram, TikTok, and YouTube have learned about some women who deserve to be better known — Corita Kent, Katherine Johnson, Hetty Green, Rosalind Franklin and (soon) Stephanie Shirley. AI does some heavy lifting, but for now, it still takes a human to write a story that a human wants to read.