The Chef and the Overeager Sous-Chef: Learning to Co-Create with LLMs
By Anthea Roberts and Nick Lothian
Picture yourself as a chef with an overeager sous-chef. You ask for a dish with a touch of thyme or a bit more lemon, and they whip it up in seconds—but there's always a twist. The thyme is overwhelming, or the lemon's barely there. You keep going back and forth, each revision bringing some improvements, yet somehow the results are never quite right.
This is what co-writing with a large language model (LLM) often feels like: an endless back-and-forth of revisions where the AI’s enthusiasm doesn’t quite match your creative intent. This has made us wonder: how can we learn to co-create better with these silicon chefs?
The Promise and Perils of Working with LLMs
At first, working with an LLM seems like a dream come true. You, the human, are the head chef—prompting, shaping, directing the creativity. The LLM, like that sous-chef, is fast, responsive, and brimming with content. Need a quick draft? It’s done in seconds. Want an alternative phrasing? It’s ready before you’ve finished asking. The AI brings ideas you hadn’t even considered, weaving in unexpected flourishes of prose, like a sous-chef adding a dash of saffron where you’d only imagined salt.
The tyranny of too many options looms large when the LLM churns out draft after draft, making it hard to avoid the feeling that you’re circling the right target but never quite hitting the bull’s eye. Editing starts to feel less like crafting a masterpiece and more like chasing a mirage. And just as an overzealous sous-chef might overdo certain elements, LLMs, without proper guidance, can overwhelm you with words, leaving you paralyzed by choices rather than inspired by clarity.
Yet despite the frustration, there’s undeniable potential here. These digital sous-chefs can whip up options and offer twists that inspire entirely new ideas. Sometimes, the sous-chef gets it right, adding a flourish that elevates the dish—or the draft—to a level you didn’t anticipate. Those moments capture the potential of co-writing with LLMs: they push the boundaries of what you might have done alone. The challenge lies in learning how to harness that potential without being overwhelmed.
The Evolution of Creative Tools: From Frustration to Precision
Where the process has generally faltered has been in the experience of editing. We’re used to working with tools we control—tweak a sentence, adjust a paragraph, done. But LLMs typically operate differently. Ask them to "bring out the lemon," and instead of a subtle tweak, they rewrite half the dish—leaving you with more new choices rather than refining the original. It’s like asking the sous-chef to add a touch of salt, only to find they’ve decided to reinvent the entire course. The problem is that AI can typically only generate so far, rather than edit effectively.
What we need are tools designed specifically for iterative, collaborative writing—interfaces that allow for precise edits without triggering wholesale rewrites, and that let you, the head chef, stay in control of the creative vision. This is what is possible with Claude’s artifacts or GPT-4’s new canvas edition (see next image). Drafting is done collaboratively, and edits can be more easily targeted. You can effectively ask the LLM to “bring out the lemon without overpowering the thyme,” transforming the sous-chef from an overenthusiastic helper into a trusted collaborator.
Just like Figma allows designers to iterate without losing sight of the whole, or DAWs let music producers refine a single note, writing with LLMs should let us focus on a phrase or sentence without reworking everything—preserving the best parts of earlier versions while seamlessly integrating new ideas. We should be able to make line-item (and big picture) suggestions to LLMs and LLMs should be able to make line-item (and big picture) suggestions to us.
Notion’s “Write with AI” offers a glimpse of what targeted LLM-assisted editing can achieve, similar to the GitHub Copilot experience that software developers are now accustomed to.
Google’s NotebookLM takes this further by generating context-sensitive suggestions tailored to what you are working on. For example, it can propose questions based on a product roadmap guide you’ve uploaded, providing highly relevant, in-the-moment insights.
The Future Kitchen: AI as Co-Head Chef
This evolution is not just a matter of improved tools; it requires a shift in our mindset. As writers, we need to rethink what it means to truly collaborate with a machine, moving beyond merely using it as a tool. Co-writing with an LLM is about more than simply directing content—it’s about learning to balance control with collaboration, accepting that sometimes the AI’s unpredictability will spark inspiration while recognizing that our human intent should remain at the core of the work.
This, ultimately, raises a larger question: what will authorship mean when machines become true partners in the creative process—not just tools for revision, but co-authors in their own right?
In the kitchen of the future, the sous-chef will likely evolve into a co-head chef, better understanding your creative vision and anticipating your needs. AI has potential not only to assist but to enhance your craft, allowing for a true partnership in creation. The creative process will hopefully feel less like wrestling with a buffet of almost-right drafts and more like two chefs working together, each contributing their strengths to produce something greater than either could achieve alone.