AI as collaborator, not mind reader

The same pattern shows up in every domain where I've deployed AI tools. Engineers frustrated that the coding assistant writes unfamiliar patterns. Executives frustrated that the drafting tool produces generic prose. Product managers annoyed that the summary tool emphasises the wrong parts. In each case, the model has been asked to produce output in someone's style without being shown that style. The output reverts to the default. The default is by definition nobody in particular, which is why it feels generic.

What 'teach it your intent' actually means

'Train a model on your own code or writing' is the phrase I've been using, and it's worth being precise about what that actually requires operationally, because the word 'train' is doing several jobs at once. At the fine-tuning end, it means producing a custom model variant with weights adjusted to your specific corpus. That's expensive, fragile, and almost never the right answer for individuals or small teams. Most fine-tuning projects I've watched produce worse results than the same effort spent on prompt and context engineering would have.

At the context-engineering end, it means something much lighter. A persistent system prompt with your preferences, principles, and style guidelines. A set of examples from your prior work, stored somewhere the model can retrieve them. A short description of the voice you want. A few dos and don'ts. This is boring engineering work. It takes an afternoon to set up. It produces dramatically better alignment than out-of-the-box generation. It is also, in the research I've seen, responsible for roughly 95% of the 'model sounds like me' experience.

A practical structure

The structure that has worked, for me and for the executives I've advised, is a personal system prompt of about five hundred words that covers: what the person writes about most often, the voice and register they prefer, the specific things they avoid (jargon, clichés, particular phrases), a list of influences whose style they want to echo, and three to five short examples of their best past work. The examples are the load-bearing part. Without them, the prompt describes a voice abstractly. With them, the model can infer the patterns directly.

The iteration that follows takes about a month. The first version of the prompt produces output that's recognisably closer to the person, but not quite right. You read the output, notice what's still off, and edit the prompt. After four or five such cycles, the output is close enough that you're spending less time rewriting than you were writing from scratch. That breakeven point is the interesting one. It arrives later than people expect and then shifts the cost-benefit of the whole workflow.

The session-persistence problem

The collaborator framing in the original post is useful, but it's worth caveating one specific limitation. A human collaborator accumulates memory of your preferences across months and years. Current AI tools reset on session boundaries. This means 'teaching the model your style' is not a one-time event that compounds - it's a configuration state you have to engineer to persist. Tools vary in how well they handle this. Some persist memory across sessions within an interface. Others require you to re-inject context every time. Knowing which category your preferred tool is in matters a lot for whether the teaching you did yesterday affects the output today.

This is changing. The direction of travel across all major AI products is toward more persistent user-level memory, longer context windows, and better retrieval of past conversations. But the specific experience of 'I taught the model my style and it forgot' is a real failure mode of current tooling, and it's worth being aware of when you're evaluating whether the collaboration is stable or just aspirational.

The examples are the load-bearing part. Without them, the prompt describes a voice abstractly.

The privacy trade-off

The other caveat that deserves mention is that 'teach it your intent' typically means uploading your past work to a third-party service. For casual personal writing, this is usually fine. For professionals whose work is sensitive - lawyers, doctors, advisers, anyone covered by confidentiality agreements - the privacy surface of the teaching is non-trivial. The collaboration framing can make this feel less fraught than it is, and the practical answer for regulated professionals is usually a self-hosted or enterprise-deployed model with contractual guarantees about retention. Casual users should still think about what's in the examples they upload. Some mine to avoid, some to anonymise, some to include.

When the model stops being a tool

The phrase I've been using in the original post is that the model starts being a collaborator when it mirrors your thinking patterns. This is directionally right and worth making more precise. A tool does what you tell it. A collaborator brings something to the partnership you don't bring yourself. Current AI tools, configured well with your context, are somewhere between the two - they execute what you'd tell them to, and also surface associations, examples, and phrasings you wouldn't have produced alone. The more context you feed in, the more the model moves toward the collaborator end of the spectrum.

What it can't yet do - and this is the most honest limitation - is push back on you the way a human collaborator would. It won't tell you your argument contradicts something you said a month ago. It won't suggest you're being lazy on a specific paragraph. It won't say 'this isn't your best work'. These are the things a senior editor does for a writer, or a thoughtful manager does for a direct report, and they're the parts the collaboration framing over-promises on.

The practical recommendation

Spend an afternoon writing a personal system prompt with five examples of your best work. Iterate on it weekly for a month. Treat the model's output as the first draft rather than the final output. Rewrite aggressively. Over three or four months, the output will become dramatically closer to your voice, and the rewriting load will drop accordingly. This is boring and unglamorous advice that produces outsized returns, which is why I've repeated it in five different formats in the last year.

The collaboration is real. It requires work. The work is mostly context engineering, not prompt engineering, and the people who invest in it are the ones who get the genuine 'this mirrors my thinking' experience. The ones who don't invest in it are the ones who keep complaining that the model is generic. Both groups are using the same underlying capability. The difference is whether they've done the teaching.