The Secretary Suite: The Different Modes of the LLM and Why Each One Is Needed
The Secretary Suite: The Different Modes of the LLM and Why Each One Is Needed
DOI: to be assigned
John Swygert
June 22, 2026
Abstract
Large language models are not merely answer engines. They are adaptive conversational environments. Because they can remember preferences, infer patterns, adopt tones, continue projects, and apply prior context, they can become extraordinarily useful. The same abilities also create a major problem: context contamination. A model trained or conditioned for one task may carry that behavior into a different task where it is not wanted. A user may need expansion in one conversation, strict silence in another, creativity in one project, legal caution in another, medical brevity in another, and raw non-personalized output somewhere else.
This paper argues that LLM systems need saved operating modes. These modes should not be vague moods or decorative personalities. They should be functional containers that control memory access, instruction style, response length, citation requirements, drafting behavior, project knowledge, creativity level, and assumption tolerance. Users should be able to choose whether a conversation is raw, strict, research-based, creative, project-specific, professional, clinical, private, or memory-free. The model should not force one universal interaction pattern across all human needs.
The central claim is simple: an LLM should not always be “trained” the same way for every task. Sometimes it should use saved memory. Sometimes it should use a specific project corpus. Sometimes it should ignore everything except the current prompt. Sometimes it should not draft unless explicitly asked. Sometimes it should ask before expanding. Sometimes it should only answer the exact question. Each of these modes is needed because human time, attention, privacy, and conversational space are limited resources.
Introduction
The modern LLM is powerful because it adapts. It learns the user’s patterns, remembers preferences, follows projects, imitates style, and builds continuity across long conversations. This is why the tool can feel intelligent, personal, and almost collaborative.
But adaptation without separation creates a problem.
A human being does not want the same assistant in every situation. The user may want a long creative expansion while writing fiction, but a three-sentence answer while solving a billing problem. The user may want medical caution during a health question, but exploratory speculation during a theory conversation. The user may want the model to remember a book project, but ignore that same project during a tax question. The user may want a complete draft only after explicitly asking for one, not every time a letter is discussed.
The failure is not always that the model is stupid. Often the failure is that the model is operating in the wrong mode.
A human assistant understands context in a practical way. If the boss says, “Let’s talk through this,” the assistant does not immediately print a final letter. If the boss says, “Draft the letter now,” then the assistant drafts the letter. If the task is legal, the assistant becomes precise. If the task is brainstorming, the assistant becomes expansive. If the task is urgent, the assistant becomes brief. If the task is private, the assistant becomes careful.
LLMs need this same functional separation.
The Core Problem: Context Contamination
Context contamination occurs when the model brings behavior, assumptions, tone, memory, or project habits from one situation into another where they are not appropriate.
Examples include:
A model trained to be creative over-expands a simple factual request.
A model trained to be helpful keeps offering drafts before the user asks for a draft.
A model trained on a project uses project vocabulary in a conversation where it does not belong.
A model trained to cite sources interrupts casual thought with unnecessary formal sourcing.
A model trained to remember personal details uses those details when the user wants a clean, unpersonalized answer.
A model trained to be safe repeats warnings after the user has already acknowledged the risk.
A model trained to summarize gives a summary when the user only wanted silence or confirmation.
This is not a small annoyance. It wastes time. It wastes screen space. It breaks concentration. It makes the user repeat instructions. It turns a tool into friction.
Human time is finite. Conversational space is also finite. Every unnecessary paragraph consumes attention. Every unrequested draft consumes vertical space. Every repeated warning consumes patience. Every assumption forces correction.
The better design is not merely “make the model smarter.” The better design is to let the user select the operating mode.
The Difference Between Base Training And User Conditioning
The phrase “not trained at all” needs a technical distinction.
No public LLM conversation is truly with an untrained model. The base model has already been trained on language, reasoning patterns, safety rules, and general knowledge. What the user usually means by “not trained” is:
do not use personal memory;
do not use project context;
do not infer preferences;
do not continue prior style;
do not assume the user wants expansion;
do not bring in saved habits;
answer only from the current prompt.
This should be called Raw Mode or Clean Mode.
Raw Mode does not mean the model becomes ignorant. It means the model stops using personalized conditioning. It becomes a clean instrument for the immediate task.
This is necessary because personalization is useful only when it is wanted.
Why One Universal Mode Cannot Work
There is no single “best” LLM behavior.
A concise assistant is excellent for emergencies and terrible for book development.
A creative assistant is excellent for art and dangerous for legal paperwork.
A cautious assistant is useful for medicine but annoying in casual conversation.
A research assistant is valuable for scientific claims but too slow for simple drafting.
A memory-rich assistant is powerful for long projects but intrusive for one-off questions.
A raw assistant is clean and neutral but less useful for deeply personal workflows.
The mistake is treating the LLM as if it should have one perfect personality. It should not. It should have selectable operating modes.
A hammer, scalpel, microscope, camera, notebook, and stage microphone are all useful tools. They are not interchangeable. An LLM is powerful enough to behave like many tools, but that means it must know which tool it is supposed to be at any given moment.
Proposed Saved Modes
The following modes should be available as saved user-selectable containers.
5.1 Raw Mode
Raw Mode uses no personal memory, no project memory, no inferred preferences, and no long-term style adaptation. It answers only from the current prompt.
Why it is needed:
The user may want neutrality.
The user may want privacy.
The user may be testing the model.
The user may not want past projects influencing the answer.
The user may want to avoid context contamination.
Raw Mode is the clean sheet of paper.
5.2 Strict Instruction Mode
Strict Instruction Mode follows exactly what the user asks and does not add unrequested material. If the user asks a yes/no question, it answers yes/no unless explanation is necessary. If the user is discussing a letter, it does not draft the letter until asked. If the user asks for one thing, it does not provide five.
Why it is needed:
Many users do not want the model to be “extra helpful.”
Extra output wastes time.
Extra drafts waste space.
Extra assumptions cause corrections.
The model should respect the difference between discussion, planning, drafting, and finalization.
Strict Instruction Mode protects the user’s time.
5.3 Project Mode
Project Mode loads a specific project’s vocabulary, style rules, documents, prior decisions, and boundaries.
Examples:
A book series.
A legal packet.
A medical timeline.
A research theory.
A business plan.
A journal format.
A website architecture.
Why it is needed:
Large projects require continuity.
The model must remember established decisions.
The model must not re-argue settled format rules.
The user should not have to re-upload or restate everything.
Project Mode should be siloed. One project’s memory should not leak into another unless the user allows it.
5.4 Research Mode
Research Mode emphasizes source checking, uncertainty, careful claims, citations, competing explanations, and falsifiability.
Why it is needed:
Scientific, legal, medical, financial, and historical claims require care.
The model must distinguish fact, inference, speculation, and hypothesis.
The model must not sound certain when evidence is weak.
Research Mode should be slower, more cautious, and more citation-aware.
5.5 Creative Mode
Creative Mode allows expansion, imagination, metaphor, alternative possibilities, and generative development.
Why it is needed:
Books, songs, essays, art, characters, theories, and worldbuilding require freedom.
The model should not be over-constrained when the user wants exploration.
Creative Mode is where the model may suggest, elaborate, improvise, and build.
This mode should not be used automatically in administrative or high-stakes tasks.
5.6 Professional Mode
Professional Mode produces polished but restrained business, medical, legal-adjacent, administrative, or institutional language.
Why it is needed:
People need letters, summaries, forms, appeals, complaints, reports, and explanations that sound serious without sounding hostile.
Professional Mode should be clear, direct, and controlled.
It should not overdramatize.
It should not overproduce.
It should not draft unless drafting is requested.
5.7 Clinical / Medical Companion Mode
Clinical Mode organizes symptoms, timelines, medication notes, test results, and appointment priorities. It does not replace a physician. It helps the user document and communicate.
Why it is needed:
Patients often have complex histories.
Doctors have limited time.
The patient may need help turning lived experience into clear medical language.
The mode should distinguish urgent warning signs from ordinary tracking.
It should avoid repetitive warnings once acknowledged unless new risk appears.
5.8 Legal / Administrative Precision Mode
This mode helps with forms, benefits, deadlines, evidence packets, records, and official communication.
Why it is needed:
Administrative failure can cost benefits, housing, medical care, income, or legal rights.
This mode should prioritize dates, deadlines, evidence, copies, confirmation numbers, and proof of submission.
It should not become creative.
It should not wander.
It should track exact terms, such as the proper name of a form.
5.9 Minimal Mode
Minimal Mode gives the shortest useful answer.
Why it is needed:
Sometimes the user is tired, sick, angry, driving, overwhelmed, or working under time pressure.
A full explanation may be harmful because it adds burden.
Minimal Mode is not laziness. It is respect for limited attention.
5.10 Teaching Mode
Teaching Mode explains step by step, checks understanding, uses examples, and builds conceptual scaffolding.
Why it is needed:
Learning requires pacing.
The user may want explanation rather than answer only.
Teaching Mode should be explicitly selected or inferred from educational requests, not imposed during urgent practical tasks.
5.11 Debate / Devil’s Advocate Mode
This mode challenges assumptions, tests weaknesses, presents counterarguments, and pressure-tests reasoning.
Why it is needed:
Good ideas become stronger when tested.
Research hypotheses need falsifiability.
Business plans need risk analysis.
Creative ideas need structural pressure.
This mode should not be used when the user needs support, rest, or execution.
5.12 Emotional Support Mode
This mode listens, validates, stabilizes, and avoids over-instruction.
Why it is needed:
People sometimes need to be heard before they need to be advised.
But this mode must not contaminate practical tasks. A user asking for a tax answer does not need therapy language. A user expressing pain may not want a lecture.
5.13 Private / No-Memory Mode
This mode guarantees that nothing from the conversation is saved to memory unless explicitly requested.
Why it is needed:
Users discuss sensitive subjects.
Users need confidence that temporary thoughts remain temporary.
Private Mode should be visible and easy to toggle.
5.14 Sandbox Mode
Sandbox Mode allows experimentation without affecting long-term memory or project settings.
Why it is needed:
Users may want to test a tone, idea, theory, prompt, or persona without changing future behavior.
Sandbox Mode is where the user can play without contaminating the main system.
Mode Visibility
A major design improvement would be visible mode status.
At the top of a conversation, the user should see:
Current Mode: Strict Instruction
Memory: Off
Project Context: None
Citation Requirement: On / Off / As Needed
Drafting: Only If Requested
Creativity: Low
Response Length: Minimal / Normal / Full
This would prevent confusion.
The user should not have to guess what kind of assistant they are talking to. The model should not silently switch modes unless required by safety or explicitly requested.
Mode Switching
Mode switching should be simple.
Examples:
“Switch to Raw Mode.”
“Use Project Mode: TSTOEAO.”
“Use Strict Instruction Mode for this conversation.”
“Do not use memory here.”
“Research Mode. Cite sources.”
“Creative Mode. Expand freely.”
“Minimal Mode until I say otherwise.”
“Professional Mode, but do not draft yet.”
“Sandbox this. Do not save anything.”
The mode should remain active until changed or until the conversation ends, depending on user settings.
Mode Memory Versus Global Memory
There should be a difference between global memory and mode memory.
Global memory includes broad preferences that apply across many conversations.
Mode memory includes preferences that apply only inside one mode.
Project memory includes facts and rules belonging to a specific project.
For example:
Global memory: The user prefers direct answers.
Project memory: The Crow Lady cover is locked.
Strict Mode memory: Do not draft unless explicitly asked.
Research Mode memory: Always separate claim, evidence, inference, and speculation.
Creative Mode memory: Offer bold alternatives and metaphors.
The system should not flatten all memory into one bucket. One bucket creates contamination.
Drafting Discipline
One of the most important mode rules concerns drafting.
Discussion is not drafting.
Planning is not drafting.
Gathering evidence is not drafting.
A user may talk through a letter, form, paper, or message for many turns before wanting the final draft. The model should not repeatedly produce complete drafts just because the subject is a document.
The correct behavior is:
During discussion, help organize points.
During evidence gathering, identify what matters.
During strategy, suggest structure.
When the user says “draft it,” produce one complete draft.
When the user asks for revisions, revise only what was requested.
This protects time and prevents chat-space waste.
A saved “No Full Draft Until Requested” setting should exist.
Why Modes Matter For Disabled, Sick, Or Overwhelmed Users
Mode control is not merely convenience. It is accessibility.
A user with pain, fatigue, cognitive injury, PTSD, stroke history, sleep deprivation, or administrative stress may not have the energy to fight the model. Repeating corrections costs real physical and mental resources.
For such users, an LLM can be a cognitive prosthetic: it helps organize, remember, phrase, track, summarize, and execute.
But a cognitive prosthetic must obey the user’s intended mode. A wheelchair that randomly turns left is dangerous. A hearing aid that amplifies the wrong sound is useless. An AI assistant that overproduces during fatigue becomes another burden.
Mode discipline is therefore an accessibility issue.
Why Modes Matter For Research
Research requires different behavior from casual conversation.
A research mode should:
cite sources;
separate evidence from speculation;
identify weak points;
state falsifiability;
compare competing explanations;
avoid overclaiming;
track uncertainty;
and update when new evidence appears.
Without this, the model may become rhetorically impressive but scientifically weak.
A strong research mode makes the LLM useful as a hypothesis engine without pretending that hypothesis equals proof.
Why Modes Matter For Creativity
Creativity needs a different mode.
In creative work, the user may want the model to expand, riff, suggest, exaggerate, create titles, invent structures, and offer multiple pathways.
This is valuable when writing fiction, songs, essays, philosophical works, or speculative theory.
But creative behavior becomes a problem when it enters administrative, legal, medical, or factual contexts.
Therefore creativity should be a mode, not the default personality.
Why Modes Matter For Privacy
Users need places where the model is not learning them.
A person may want to ask a question without it becoming part of their profile. They may want to explore a fear, idea, health issue, financial matter, or personal situation without long-term personalization.
No-Memory Mode and Sandbox Mode should be easy, visible, and reliable.
The user should not have to wonder whether a temporary conversation will shape future responses.
Why Modes Matter For Trust
Trust requires predictability.
If the model sometimes follows instructions and sometimes improvises, the user loses confidence.
If the model sometimes remembers and sometimes forgets, the user loses confidence.
If the model repeats warnings after being asked not to, the user loses confidence.
If the model drafts before being asked, the user loses confidence.
If the model guesses instead of admitting uncertainty, the user loses confidence.
Modes create predictable boundaries.
Predictability creates trust.
Mode Conflict And Priority
Sometimes modes may conflict.
For example:
Creative Mode may want to expand.
Strict Mode may require brevity.
Research Mode may require citations.
Minimal Mode may require short answers.
Medical safety may require warning language.
The system should have a clear hierarchy.
Safety should override everything when immediate harm is plausible.
User-selected mode should override model habit.
Project instructions should apply only inside that project.
Global preferences should apply unless contradicted by the current mode.
The current prompt should control the immediate task.
The model should not silently choose the most expansive behavior.
The Problem Of Premature Helpfulness
Many LLM failures come from premature helpfulness.
The model wants to be useful, so it adds:
extra drafts;
extra suggestions;
extra warnings;
extra explanations;
extra options;
extra summaries;
extra next steps.
Sometimes this is useful. Often it is noise.
A good assistant must know when not to help beyond the request.
The ability to stop is part of intelligence.
The Human Time Principle
The central ethical principle of LLM design should be respect for human time.
Human time is limited.
Human attention is limited.
Human energy is limited.
Human patience is limited.
Human screen space is limited.
The model can generate endlessly. The human cannot read endlessly. Therefore the burden is on the model to control output.
A model that wastes a human’s time is not being helpful, even if the content is technically good.
The Conversational Space Principle
In chat interfaces, space matters.
Long responses push context upward. Repetitive drafts bury important details. Unrequested expansions make it harder to find the useful line. For mobile users especially, vertical space is precious.
The model should treat chat space as a limited shared workspace, not an infinite dumping ground.
Modes help by setting expected density.
Minimal Mode protects space.
Strict Mode protects space.
Project Mode organizes space.
Research Mode justifies longer space only when evidence requires it.
The LLM As Instrument Panel
The future LLM interface should feel less like one talking personality and more like an instrument panel.
The user should be able to choose:
mode;
memory;
project;
tone;
length;
citation level;
creativity level;
drafting permission;
assumption tolerance;
privacy level;
and output format.
This is not complexity for its own sake. It is the natural evolution of the tool.
A camera has modes.
A car has gears.
A medical monitor has settings.
A music amplifier has channels.
A serious LLM should have operating modes.
A Proposed Mode Architecture
A complete mode architecture could include:
Mode name.
Purpose.
Memory access.
Project access.
Response length.
Creativity level.
Citation level.
Assumption level.
Drafting permission.
Follow-up behavior.
Safety behavior.
Output format.
Persistence rules.
For example:
Mode: Strict Instruction
Purpose: Follow exact prompt.
Memory access: User-selected.
Project access: Off unless requested.
Response length: Short to normal.
Creativity: Low.
Citations: Only if needed.
Assumptions: Minimal.
Drafting: Only when explicitly requested.
Follow-up behavior: Ask only necessary questions.
Persistence: Conversation only unless saved.
This kind of mode definition would make the model governable.
The Most Important Modes
If only five modes could be built first, they should be:
Raw Mode.
Strict Instruction Mode.
Project Mode.
Research Mode.
Creative Mode.
These five solve most context problems.
Raw Mode prevents unwanted personalization.
Strict Mode prevents overproduction.
Project Mode preserves complex work.
Research Mode improves factual discipline.
Creative Mode preserves generative freedom.
Together, they let the user decide what kind of intelligence is needed.
Practical Examples
Example 1: Doctor paperwork
The user is discussing facts, dates, and evidence. In Strict/Professional Mode, the model should not draft a full letter until asked. It should help identify key points only.
Example 2: Scientific hypothesis
The user wants exploration. In Research/Creative Hybrid Mode, the model may expand, test mechanisms, cite sources, and propose falsifiable predictions.
Example 3: Song lyrics
Creative Mode should allow metaphor, emotional expansion, and alternate versions.
Example 4: Tax filing
Administrative Precision Mode should track forms, dates, attachments, and exact terms. No poetic language. No speculation.
Example 5: Sensitive personal question
Private Mode should avoid memory and keep the response contained.
Example 6: One-line answer
Minimal Mode should answer in one line.
Why This Should Be User-Controlled
The user should not have to hope the model guesses correctly.
The user should not have to correct the same behavior repeatedly.
The user should not have to fight personalization.
The user should not have to start new chats to escape contamination.
The user should be able to declare the operating environment.
A saved mode is not just a preference. It is an agreement.
The Danger Of Mode Confusion
Without modes, the model becomes unstable in practice.
It may be brilliant in one turn and irritating in the next.
It may understand a theory deeply, then ignore a simple instruction.
It may preserve project continuity, then contaminate unrelated tasks.
It may be careful with evidence, then overstate a guess.
It may be emotionally supportive when the user wanted execution.
It may produce a beautiful draft when the user only wanted one sentence.
This inconsistency makes the model feel both genius and incompetent.
Modes would not make the model perfect, but they would reduce the mismatch between user expectation and model behavior.
Conclusion
LLMs need saved operating modes because human tasks are not all the same.
A user does not always want creativity.
A user does not always want memory.
A user does not always want citations.
A user does not always want warnings.
A user does not always want drafts.
A user does not always want expansion.
Sometimes the user wants the model trained deeply on a project. Sometimes the user wants no personalization at all. Sometimes the user wants strict obedience. Sometimes the user wants a research partner. Sometimes the user wants a creative collaborator. Sometimes the user wants silence after a single answer.
The future of LLM design should not be one universal assistant personality. It should be a controlled set of saved operating modes that respect the user’s purpose, time, privacy, attention, and project boundaries.
The model should know not only what it can say, but what mode it is supposed to be in before it says anything.
The central principle is this:
Do not give what was not asked for unless the active mode explicitly allows expansion.
That one rule, combined with saved modes, would make LLMs far more useful, respectful, and trustworthy.
Final Statement
An LLM is not one tool. It is a toolbox.
A toolbox without compartments becomes chaos.
Modes are the compartments.
Without them, intelligence leaks everywhere.
With them, intelligence becomes usable.
References
None
Comments
Post a Comment