I recently experimented with using Large Language Models (LLMs) to enhance biographical writing, and I want to share my approach because it tackles some interesting challenges in working with AI. The process taught me a lot about both the capabilities and limitations of LLMs, and how to work around them effectively.
Let me start with why I chose a multi-step approach. One of the biggest constraints when working with LLMs is the context window - there's only so much information you can feed into the model at once. While newer models have larger context windows, they still have limits. You might think "Why not just give the LLM all the biographical info and historical context at once?" The answer lies in the way LLMs process and integrate information.
When you dump too much information into a single prompt, LLMs tend to either lose focus on important details or fail to deeply integrate different aspects of the information. It's like asking someone to juggle too many balls at once - something's bound to get dropped. By breaking the process into distinct steps, we can get much better results.
Here's how I approached it: I started with the raw biographical data about John Robert Smith - just the basic facts about his life. This gave the LLM a clear foundation to work with. Think of it as laying down the basic structure of a house before adding the details that make it a home. Here's what that initial input looked like:
"I, John Robert Smith, was born on February 19, 1959, in San Leandro, California, a town nestled near the San Francisco Bay. My parents, Robert and Helen Smith, were part of the Greatest Generation, having lived through the tumult of World War II. My father worked as a shipbuilder in the Oakland shipyards, while my mother was a schoolteacher."
In the second step, I asked the LLM to generate historical context for two pivotal years: 1959 and 1977. This research covered everything from the Cuban Revolution to the rise of rock and roll, from early computing to the Civil Rights Movement. By keeping this separate, the LLM could focus solely on creating rich, accurate historical context without trying to weave it into the narrative yet.
The final step was where the magic happened - integrating everything together. Let me show you what the final result looked like for just the opening section:
"I, John Robert Smith, entered the world on February 19, 1959, in San Leandro, California—a small town cradled by the shimmering waters of the San Francisco Bay. My parents, Robert and Helen Smith, were emblematic of the Greatest Generation, having weathered the storms of World War II. My father toiled as a shipbuilder in the bustling Oakland shipyards, while my mother nurtured young minds as a schoolteacher. Their post-war optimism infused our household with a sense of hope and possibility, even as the world outside grappled with its own uncertainties."
Look at how the basic facts have been transformed. The same information is there, but now it's woven into a narrative that connects personal experience with historical context. The shipyard isn't just a workplace - it's a symbol of post-war industrial America. The parents aren't just workers - they're representatives of their generation's resilience and optimism.
The limitations of LLMs actually helped shape this approach. LLMs can sometimes "hallucinate" or make up details when asked to handle too much at once. By separating the factual biographical input from the historical research and then the narrative integration, we reduce the risk of the LLM inventing false biographical details or misattributing historical events.
The other benefit of this approach is its flexibility. Want to add more historical context? Just run another focused query in step two. Need to adjust the tone of the narrative? You can tweak the integration step without touching the factual foundation. It's like having modular building blocks that you can adjust independently.
I've noticed the final product has a depth that would be hard to achieve with a single-step approach. The biography doesn't just tell us what happened in John's life - it shows us how his personal story was shaped by and connected to the broader currents of history. We see how the optimism of the post-war era influenced his family life, how the space race inspired his career choices, and how the cultural shifts of the 60s and 70s provided the backdrop for his coming of age.
Looking ahead, I think this approach has potential beyond just biographical writing. Any project that requires integrating personal narratives with broader historical or social contexts could benefit from this stepped method. It's about understanding the limitations of our tools and designing workflows that play to their strengths while mitigating their weaknesses.
What I find most exciting is how this process creates something greater than the sum of its parts. The final biography isn't just a collection of facts and historical events - it's a rich tapestry that helps us understand both an individual life and the times that shaped it.
This multi-step LLM approach is now integrated into our Lifebook product, where it operates continuously during the autobiography creation process. As authors, friends, and family members contribute their stories and memories, the system automatically enriches and contextualizes the narrative. Each new input triggers the process, growing the autobiography organically while maintaining historical accuracy and narrative cohesion.
But perhaps the most meaningful application of this technology lies in its potential to help those facing cognitive challenges. At voiceofreason.ai, we're focusing on how these detailed autobiographies can serve as memory aids for individuals with dementia. Having access to a rich, contextualized life story can help maintain connections to personal history and identity. The same LLM technology that weaves historical context into personal narratives can help make memories more accessible and meaningful for those who need them most.