A theory developed by Flux+Form. The gradual process by which a large language model’s output comes to carry the individual “flavor” of the person using it, shaped by their prompting voice, their recurring patterns, and the memory the tool keeps across conversations, such that the same prompt yields a meaningfully different result for one person than it does for another. The name draws an analogy to a cast iron skillet, whose cooking surface is built up over years of use into a one-of-a-kind environment that cannot be bought or transferred. For agencies, LLM Seasoning reframes prompting as an accumulated craft rather than a one-time instruction, and raises an unresolved strategic question about whether that craft is a defensible competitive advantage or simply a head start that copies the moment it is written down.
Also known as model seasoning, prompt patina, AI voice adaptation. Term coined by Jeremy Swiller, Flux+Form.
Most people picture a large language model as a vending machine: the same input produces the same output, every time, for everyone. This mental model is wrong in a way that matters. A language model does not retrieve a finished answer; it improvises a fresh one in direct response to how the request is made. Change the phrasing, the structure, the order of ideas, or even the sentiment underneath the request, and the model generates something different in return. The person asking is not a neutral party pressing a button. They are a variable in the output.
LLM Seasoning is the compounding version of that effect. It rests on three observable behaviors of modern models, stacked on top of one another. First, generation is improvised rather than canned, so how you ask always shapes what you get. Second, these models are notoriously sensitive to subtle variations in prompt phrasing and structure, and that sensitivity extends to formatting, sequencing, and tone, meaning a user’s individual voice measurably influences results. Third, memory features retain a running record of how a given person works and carry it across conversations, by deliberate design, so that the same question produces different answers for different people. Together these mean that the longer a person works with a tool in their own consistent voice, the more the output surface takes on their particular flavor.
The term is a Flux+Form coinage rather than an established industry standard. What is well supported is the underlying mechanism: prompting voice, sensitivity, and memory are all documented. What the term adds is a single handle for the cumulative result, and a frame for thinking about what that result is worth.
Agency output is judged on voice, taste, and craft, the exact qualities that survive in how a person prompts rather than in the prompt itself. That makes seasoning more consequential here than in fields where AI is used for routine extraction or summarization.
It explains why identical tools produce unequal results across a team. When a senior strategist gets usable output from the same model a junior calls useless, the difference is rarely the subscription tier. It is months of accumulated seasoning: a sharper sense of how to ask, a refusal to accept the first generic answer, and a memory the tool has built around that person’s standards. Recognizing this prevents agencies from blaming the tool when the real variable is the operator.
It reframes prompting as a trainable craft, not a search-box trick. If results scale with how a person works rather than with what they paste, then prompting skill is a developable asset that compounds over time. Agencies that treat it as a discipline, with shared standards and deliberate practice, build capability that agencies treating prompts as disposable one-liners never accumulate.
It surfaces an unresolved question about defensibility. If seasoning is real and personal, it may be a genuine moat, a way of working with AI that a competitor cannot acquire just by buying the same software. But the moment a way of prompting can be written down, it can be handed to anyone. Whether seasoning is an owned advantage or a copyable head start is not yet settled, and agencies betting their positioning on AI craft should hold that question honestly rather than assume the flattering answer.
A mid-sized independent agency runs a test after noticing that AI-assisted first drafts vary wildly in quality depending on who produced them. They take a single brand-voice brief and a fixed copywriting prompt, then have three people run it through the same model: a senior creative who has used the tool daily for a year with memory enabled, a junior hire two weeks in, and a logged-out incognito session with no history at all.
The senior creative’s draft lands closest to the brand voice and needs roughly fifteen minutes of editing. The junior’s draft is serviceable but generic, the kind of competent-but-flat copy the team associates with AI, and needs closer to an hour of rework to match the brand. The incognito draft, interestingly, comes back stronger than the junior’s, because the prompt itself was written by the senior creative and carries real weight on its own. That last result complicates the tidy story: it shows the recipe matters enormously, not only the seasoning.
The agency draws a practical conclusion rather than a triumphant one. They stop assuming the tool is the bottleneck, they pair junior staff with senior prompters to transfer what can be transferred, and they begin documenting the prompting patterns that travel well while accepting that some portion of the senior creative’s edge lives in a seasoned working relationship that no document fully captures. The exercise is less about proving a moat exists and more about learning which parts of their AI craft are teachable and which parts are earned.
LLM Seasoning is a frame, not a finding. The mechanisms underneath it are documented and real. The bigger claim, that the flavor you build into a tool is a durable advantage only you can wield, is genuinely too early to call. A strong enough prompt can be rendered beautifully by a cold, blank machine that knows nothing about you, which is evidence against the idea that the seasoning is the whole story. Yet that strong prompt was itself developed inside a seasoned way of working.
We are publishing the term because the question it raises is worth holding, not because we have resolved it. Anyone claiming certainty about whether AI craft is a moat is, at this stage, selling something. This entry will be updated as the evidence develops.
If how your team works with AI is an accumulating asset rather than a button they press, it can be taught, sharpened, and made consistent across the agency. The Creative Cadence Workshop is eight weeks of hands-on AI training that turns individual prompting instinct into a shared, repeatable practice your team keeps.