how to write with an LLM (without losing your voice)

Large language models are often treated as machines for producing finished text. I have found them more useful as conversational spaces in which ideas can be explored, tested and gradually stabilised. The model is not asked to replace thought but to interact with it. Fragments, intuitions, contradictions and incomplete arguments are all valid starting points because the objective is not immediate fluency but the progressive organisation of meaning.

The approach did not emerge from experimentation with AI alone. It draws on decades of work developing a relational model and conceptual scaffolding for logic, language and communication, informed by systems theory, topology, harmonic organisation and the dynamics of persistence. Recursion is central to that work: a theory of communication must eventually encounter itself as communication, just as a method for organising language must become capable of reorganising its own terms. Writing with an LLM is not the theory, but one practical application through which the theory can recursively inspect, test and refine its own expression.

The process is iterative. Rather than requesting a polished essay, begin with whatever structure already exists, however incomplete. Each exchange subtly reorganises the relationships between concepts, examples, rhythm and emphasis. As those relationships change, previously hidden connections become visible and the underlying structure grows clearer. The model proposes language; the human corrects its orientation, rejects what is wrong, preserves what is right, and the conversation gradually converges. Writing becomes less like dictation and more like tuning a resonant system until its organisation can sustain itself.

That convergence is rarely immediate. I review most texts repeatedly, often moving through many rounds of revision before the language is sufficiently precise. The apparent speed of generation can obscure how much judgement, rejection, restructuring and close reading the process still requires. In general, writing this way involves about as much work as writing manually. The labour has not disappeared; it has shifted from producing every sentence directly to directing, evaluating and repeatedly reorganising the whole.

This only works if authorship remains with the human. The model contributes variation, perspective, memory, compression and expression, but conceptual direction, judgement and responsibility remain with the writer. Used this way, an LLM becomes less a generator of text than an instrument for discovering and refining ideas already taking shape. The writing is not produced by either participant alone. It emerges through sustained interaction between them.