Most interactions with an LLM are shallow by default.
You ask a question. It gives an answer. Maybe the answer is decent, maybe it is generic, but either way the model is usually working with a limited picture of who you are, what matters to you, and how you want it to behave.
If you want more control over the model, you have to give it more context.
That is the idea behind Soul Buckets.
I came up with it while building a conversational chatbot. Working with structured context made me wonder how much of a person's style and perspective could be represented in a form a model could use.
At a systems level, Soul Buckets are a structured way to provide an LLM with a much richer identity layer so it has more to work with and more behavioral guidance to follow.
One possibility that interests me is using such a profile, alongside examples of a person's writing, to approximate aspects of their style, preferences, and perspective. That is an idea to explore and evaluate, rather than evidence that a model can recreate a person.
That is not magic. It is context architecture.
What Are Soul Buckets?
A Soul Bucket is my name for a structured identity profile: categories of context about a person's priorities, preferences, and circumstances. JSON is one way to represent it. When supplied to a model, those details can guide what the answer emphasizes and how it is expressed.
The following is an illustrative profile, not a factual biography of me. Its health details, practices, goals, and concerns are examples:
{
"identity": {
"name": "Example Founder",
"role": "AI Consultant & Founder",
"values": ["clarity", "purpose", "authenticity", "continuous_learning"]
},
"health": {
"current_goals": ["consistent sleep schedule", "daily movement"],
"challenges": ["desk-bound work hours", "high cognitive load days"]
},
"career": {
"current_focus": "scaling JedAI consulting practice",
"industry": "AI / Technology",
"stage": "founder / early growth"
},
"learning": {
"active_interests": ["AI agent architecture", "stoic philosophy", "systems thinking"],
"preferred_format": "applied examples over theory"
},
"spiritual": {
"orientation": "secular mindfulness with Buddhist influence",
"practices": ["meditation", "journaling"]
},
"vision": {
"5_year": "Build a recognized AI consulting brand with recurring enterprise clients",
"core_fear": "building something technically impressive but commercially irrelevant"
}
}
The Nine Buckets
The framework organizes identity across nine domains:
- Metadata — Core identity: name, role, values
- Health — Physical and mental state, goals, challenges
- Career — Current work, industry, ambitions
- Learning — Active interests, preferred formats, knowledge gaps
- Spiritual — Orientation, practices, meaning-making frameworks
- Lifestyle — Routines, environment, energy patterns
- Vision — Long-term goals, fears, definitions of success
- Social — Key relationships, communication style, support network
- Tech — Tools used, technical depth, platform preferences
The exact buckets can change. What matters is that the structure is rich enough to shape behavior in a meaningful way.
How It Changes the Interaction
Without Soul Buckets, asking "how should I think about my pricing strategy?" gets a generic answer about value-based pricing and market research.
With a profile like the illustrative one above, the agent has context that could shape the answer:
- You're a founder at an early growth stage
- Your core fear is building something technically impressive but commercially irrelevant
- You value clarity over complexity
- You're scaling a consulting practice with enterprise clients
That gives the agent a more specific starting point. The next question is whether the added context improves the relevance of the answer in practice; that is something to evaluate, not assume.
Why This Matters
This is not necessarily about creating a better future for human-computer interaction in some abstract sense.
It is about control.
The more useful context you provide to an LLM, the more influence you have over how it behaves.
That matters if you want:
- More personalized outputs
- Stronger consistency
- Better simulation of a person's style or point of view
- More controlled agent behavior
- Richer context for decision-making
In other words, Soul Buckets are not just about making the model feel more personal. They are about making the model more configurable.
Exploring a Person's Voice
The possibility I keep thinking about is what could happen if a model had a rich, organized picture of someone: messages, emails, preferences, values, and other context, along with examples of how they express themselves.
Could it produce responses that sounded close to that person's writing? I think that is worth exploring. The idea extends even to a person who has passed away, although a simulation would remain an approximation built from the records available. It could not establish what that person would truly think or restore their consciousness.
There are different ways to explore this. Supplying a profile as context guides an interaction without changing the model's weights. Training or fine-tuning would be a separate project with its own data and evaluation requirements.
A Personal Belief Behind the Idea
For me, this connects to a belief that God created us with patterns that shape how we think and act. I picture each person as having their own rich "JSON string": a metaphor for the context, history, and priorities that make them distinctive.
That is my philosophical interpretation of what makes us distinctive. A structured profile could capture parts of that context while leaving much of a person's life and experience unrepresented.
Choose the Context Deliberately
This also brings me back to the digital footprint. I would not start by collecting every available message simply because it exists. The useful question is which context serves the purpose, whose information it is, and whether its use is appropriate and agreed to.
A small, relevant, maintained profile can be a better starting point than a large archive of sensitive or outdated material.
That can be used for personalization, simulation, creative projects, or more controlled agent design.