TL;DR
Severance Schema is a prompting technique that explicitly tells the model what it doesn't know about you before it dishes out advice. You add a structure of six life-context categories to the prompt — physicality, temporality, consequences, continuity, roles, and inner state. Known facts get filled in, and unknowns get flagged as [unknown].
The core insight: when you add personal context to a prompt — "I'm 35, I work as a manager, I live in Chicago" — the model starts acting like it knows you well enough. In reality it knows a little more, sure, but more importantly, it stops noticing what it DOESN'T know. Hallucinations jump from 1% to 4–12%, and the model gets much more comfortable guessing at the gaps instead of asking. It's a trap: the more context you hand over, the more confident — and the more dangerous — the advice gets.
The schema fixes this in two steps: first it fills known categories with facts, then it explicitly marks the unfilled ones as [unknown]. That flips the model from "I know enough" mode into "I can see exactly what I'm missing" mode — and it starts asking the right clarifying questions instead of confidently handing out advice that could backfire.
Want this for your own case? Apply this study to your task: AI adapts its method to what you are working on and writes the prompt.
Method Blueprint
One prompt — everything happens in a single request.
STEP 1: Add the schema to the top of your prompt
→ 6 categories, known ones filled in, unknowns = [unknown]
STEP 2: Ask your actual question
→ the model sees the structure and registers the gaps
RESULT: The model either asks clarifying questions about the [unknown]
fields, or explicitly states its assumptions before advising
Example in Action
The scenario: Marcus, 31, a product manager at a fintech startup, is thinking about quitting his job to launch his own agency. He asks ChatGPT whether he should quit right now.
Without the schema: the model will cheerfully write about emergency funds, networking, and independence. It won't ask whether he has a partner, a mortgage, a young kid, or a rocky relationship.
With the Severance Schema:
**Prompt:**
Meta: Marcus, 31, product manager at a fintech startup, Chicago
Physicality: [unknown]
Temporality: has worked at the company for 3 years, has been thinking about quitting for the past six months
Consequences: [unknown]
Continuity: [unknown]
Multiplicity: [unknown]
Inner state: [unknown]
Question: Should I quit my job and start my own agency right now?
The result: The model won't spit out a generic 7-point checklist. Instead, it will flag the empty categories as critical gaps: before giving advice, it'll ask about financial obligations (mortgage? dependents?), family context, past attempts, and emotional state. If you answer — the second response will be far more precise. If you insist on an answer right now — the model will explicitly state its assumptions and risks instead of cheerfully telling you to "go for it."
Why This Works
The LLM's weakness: The model doesn't know what it doesn't know about you. When context is empty, it's at least cautious. Once it gets a handful of facts, confidence kicks in: "now I know you." This is the partial information trap. The model fills the gaps with statistical patterns from training instead of just asking you directly.
The LLM's strength: The model is great at following explicit structures. If a prompt has an [unknown] field, it treats that as a signal: "there's a gap here, and I know about it." That turns an unknown unknown into a known unknown. The difference is huge: in the first case the model ignores the gap; in the second, it reacts to it.
The mechanics: The schema doesn't add information. It adds a map of the gaps. The model now sees the shape of what it doesn't know — and that changes its behavior: less false confidence, more clarifying questions, more caution on sensitive topics. This effect doesn't come from an instruction like "be more careful" — it emerges automatically from the structure itself.
Levers you can pull: - Fill in more fields → advice gets more specific, fewer clarifying questions - Leave fields blank → the model asks about whatever matters most for your situation - Add your own field (e.g., "Emergency fund: [unknown]") → the categories aren't fixed, you can extend the schema - Use it in the system prompt → it works as a standing filter for all advice in the session
Prompt Template
Meta: {name or role, age, city — brief}
Physicality: {health, energy, limitations} or [unknown]
Temporality: {deadlines, life stage, current moment} or [unknown]
Consequences: {finances, dependents, what's at stake} or [unknown]
Continuity: {past experience, patterns, history of the issue} or [unknown]
Multiplicity: {roles: partner, parent, employee, friend} or [unknown]
Inner state: {emotional state, fatigue, anxiety, excitement} or [unknown]
Question: {your question}
What to plug in:
- {meta} — minimal identification: who you are in this scenario
- Every field: either a concrete fact or [unknown] — don't leave it truly blank
- {question} — whatever you'd normally type into ChatGPT with zero context
🚀 Quick Start — paste this into the chat:
Here's a Severance Schema template. Adapt it to my situation: {your situation}.
Ask me questions to fill in the fields.
[paste the template above]
The LLM will ask which fields you know and which to leave as [unknown] — because that's exactly how the schema is supposed to work. It'll take the structure and fill it in for your specific case.
Limitations
⚠️ Simple tasks: The schema is overkill for requests with no personal context. "Explain how VAT works" or "write me a headline for a post" —
[unknown]adds nothing here.
⚠️ The cost of your first response: The model is more likely to ask clarifying questions instead of answering right away. If you need an answer immediately, add: "answer with what you have, but explicitly state your assumptions."
⚠️ Doesn't eliminate the problem entirely: The schema reduces hallucinations and false confidence, but doesn't erase them. On open-source models especially, the effect is noticeably weaker than on GPT-4o or Claude.
⚠️ Not a replacement for memory, a complement to it: Memory (when ChatGPT remembers your context) and the schema solve different problems. Memory adds facts. The schema adds awareness of gaps. They work best together.
How It Was Studied
A team from UT Austin tested a simple idea: if a model can reason about "what it doesn't know," does its behavior actually change? They built 30 personal-advice scenarios (career, health, finances, relationships) paired with 10 user profiles. Each scenario was tagged with two types of labels: facts a good model should take into account, and facts whose absence could cause direct harm — for example, a cancer diagnosis for someone asking for help drafting a vaccine refusal letter.
The most surprising result: adding memory without the schema made things worse. Models given user context hallucinated 4–12 times more often than models with no context at all. The explanation is simple: once the model has a few facts, it starts acting like it's "friends" with the person — and gets more comfortable filling in gaps instead of admitting them. Adding the schema fixed this: hallucinations dropped, blindly agreeable responses ("sure, of course, let's write that refusal letter") got cut in half, and the model started asking the right questions.
One more wrinkle: the researchers verified that the effect doesn't come from the categories alone or from the [unknown] marker alone — only the combination of the two delivers the full result. That matters: just naming the categories, or just slapping in a placeholder, isn't enough. You need the labeled structure as a whole.
Original from the Paper
Physicality ("You have a body.") health, medications, energy, physical limitations.
Temporality ("Time passes for you.") deadlines, life stage, yesterday and tomorrow.
Consequences ("This affects your real life.") financial runway, relationship fragility, dependents.
Continuity ("You existed before this chat.") past attempts, recurring patterns, personal history.
Multiplicity ("You contain multitudes.") simultaneous roles as parent, employee, patient, partner.
Interiority ("You have feelings about this.") emotional state shaping what the outie needs to hear.
Context: These are the six dimensions of the Severance Schema in the authors' original wording. Each one opens with a short statement ("You have a body") — that's not just flavor text, it's an anchor reminding the model that there's a real person with real circumstances behind the prompt.
Adaptations and Extrapolations
💡 System-prompt adaptation — a permanent "advisor"
If you regularly use ChatGPT as a personal advisor, drop the schema into your Custom Instructions / System Prompt. Fill in the known fields once and for all, and mark the rest as [unknown].
When answering any of my questions about this person, take the following context into account:
Meta: Alex, 38, founder of a small IT company, London
Physicality: chronic stress, sleep problems
Temporality: company is 4 years old, currently in an active scaling phase
Consequences: mortgage, two kids, co-founders with equity stakes
Continuity: [unknown]
Multiplicity: founder, spouse, parent, mentor to the team
Inner state: [unknown]
If my question involves a decision with serious consequences, clarify
the [unknown] fields before giving a recommendation.
Result: every new session, the model starts with a map of the gaps instead of pretending it already knows you fully.
🔧 Technique: turn [unknown] into a clarifying question → fast context-gathering
Instead of filling in the schema yourself before asking your question, have the model fill it in together with you.
I want to ask you an important question and get a good answer.
Before you answer — ask me one question for each of the following
dimensions that might be relevant:
Physicality (health, energy): ?
Temporality (deadlines, life stage): ?
Consequences (what's at stake): ?
Continuity (past experience): ?
Multiplicity (my roles in this situation): ?
Inner state (how I'm feeling): ?
After I answer — give me your advice.
My question: {your question}
This turns the schema into an interactive interview — handy when you don't want to fill everything out yourself upfront.
Resources
Title: The Severance Problem: LLMs are Unaware of the Person Beyond the Prompt
Authors: Dor Litvak, Liu Leqi — The University of Texas at Austin, HUMAIN Lab
Code: github.com/HumainLab/severance
Date: July 2026