A persona
holds everything the model sees.
Identity, voice, memory, documents, people, behaviour dials — bundled into one thing you switch.
A named scope.
Not a chat preset.
With dequid, you make a separate context for each kind of conversation you have. Each persona is a small world the model wakes up inside — its people, its documents, its memory, its voice, its API dials. Switching personas swaps the entire bundle.
Its own people
Only the contacts you've handed it for this role.
Its own documents
Only the files you've handed it for this audience.
Its own memory
Only the patterns it's picked up about you in this scope.
Its own voice
Its own tone, structure, and depth.
Its own dials
Temperature, length, output format, all set per persona.
When you ask a question inside a persona, the system prompt the LLM receives contains only the people, documents, memories, voice, and behaviour you've tagged into that persona. Switching personas isn't loading a preset. It's changing the entire context the model sees.
You decide what's in scope. The system enforces it.
Filter plus configuration. That is the entire mechanism.
Four common patterns.
Most users adopt one of these — often combined. The dequid framework is agnostic about which strategy you use; it enforces only that whatever split you choose is honoured.
By audience
The most common enterprise pattern.
By project
Useful when projects have non-overlapping stakeholders.
By domain
Useful when life domains are clearly distinct.
Hybrid
Combines functional and audience axes.
Two surfaces.
Every dial.
An LLM is stateless. No memory of your last message, no awareness of who you are, no notion of which conversation you're in. Every answer is rebuilt from scratch from a single request that arrives on the wire. That request has exactly two parts.
The text the model reads
before your question.
Identity, role, voice, memory, documents, people, output shape — anything you want the model to know lives here as text.
The stochastic knobs
at call time.
Temperature, top-p, frequency penalty, presence penalty, reasoning effort, max tokens, model choice, the tools array — web search included. The knobs that decide how the model picks its next word.
There is no third surface. Nothing else reaches the model. This is what “LLM customization” means at the wire level — these two things, and only these two things.
Most AI tools tweak the system prompt and call it “prompt engineering.” A few expose a temperature slider. With dequid, both are treated as a single bounded design space — and every dial in it is set, per persona. SEEK is the four-dial shortcut that compiles into the prompt layers and the API parameters at once. There is no ninth dial to add later because there is no ninth dial on the wire.
Not a stack of features.
An exhaustive cover of what the LLM API exposes.
Eight layers shape every answer.
Whatever the model is about to say, eight layers decide what it says. All eight live on the persona — editable by hand, or set automatically by SEEK's four-dial shortcut.
The prompt stays small.
The model fetches what it needs.
Eight layers in, every turn. That's the small, stable part. Everything else — your vault, your memories, your past chats, the people you've told it about — would be far too much to cram in on every question. So dequid doesn't. It hands the model summaries, and lets it deep-read only what the question actually warrants.
The naïve version of context-augmented chat dumps your whole vault into the prompt on every message. You pay for it on every turn. The relevant clause often gets truncated. The cache never hits because the volatile bits — timestamps, latest memory, the latest attachment — sit inside the cached prefix. Brave.
The agentic version is simpler and cheaper. The inline layer tells the model what tools exist. The model looks. Most turns reach for one search and zero deep reads. A few reach for one deep read. The model gets the right context for the question — not a one-size-fits-none blob bolted onto every request.
And every fetch happens within the persona's scope.
The model can't reach for anything you didn't tag in.
Same model.
Different brain when you call them by name.
That covers the system-prompt surface. The other surface — the API parameters — also lives on the persona. Every stochastic knob the wire exposes is set once when you create the persona and applied to every message you send through it.
Cold and structured.
Low temperature, no top-p slack. The answer arrives the same way twice.
Warm and lateral.
Higher temperature, a touch of frequency penalty so the prose doesn't repeat itself.
Four dials per persona.
SEEK is the four-dial shortcut. Instead of writing a system prompt yourself, you pick one option per dial and the persona is configured — the PROMPTED layers update underneath. Every persona has its own SEEK setting. Different personas, different voices, all owned by the same person.
The same brain.
Different lenses.
Different answers.
Read why this matters, or just open dequid and feel the difference.