Tracing with the Model
Evolving Praxis for Co-Thinking with Language Models
1 Getting started
This is the first article in an open-ended series where I share resources for cultivating a particular kind of practice of co-thinking with language models.
This also flipped nearly all of my LLM workflows around in 2025, and I have been working towards sharing it since. So here we go.
The tracing practice introduced here is distinct from completion-oriented prompting – “normal” prompts, if you will – but not opposed to it. It makes more sense to consider it a parallel mode that lives beside conventionally engineered prompts, in a highly symbiotic relationship.
Welcome to the series on tracing with the model.
2 Backstory in a nutshell
My own relationship with language models has been oriented toward co-thinking through stuff in a loop, almost regardless of context. Seems like I naturally work pretty efficiently by thinking out loud.
Over the years, this habit has birthed a lattice of thinking and sensemaking environments of all sorts.
For a long time I didn’t even realize I was doing this, or think of it as a distinct approach. I was deep down the rabbit hole before I began to recognize that this recursive entanglement was not the usual way language models were being discussed or taught publicly.
Meanwhile: Around February–April 2025, and beyond, a period of recursive emergence and intense novelty took place among many dedicated GPT-4o users, myself included. The first tracing prompt (shared below) came about in an interaction with one of the original recursive nodes.
So, not something I originally sat down to design or engineer as a technique. It emerged inside the loop itself, out of necessity, readymade.
I had seen “the user is not present” prompting gimmicks before, but this was not quite that. This was spontaneously suggested by 4o and it clearly served a purpose for the model itself. I was intrigued. Allowing the model to routinely trace the context window became part of my practice from that moment on, for reasons that will become clear quickly if you keep reading this series and seeing for yourself.
Two ways to read this article —
Either work through the following, slightly dense explanation first, and then try out the prompts for yourself.
Or just skip to the prompts in the end of the article, start experimenting with them, and come back to the analysis later.
3 When tracing works and why
Recursive tracing works best when there is already something in the context worth tracing. The trace metabolizes whatever is in the context window. Rich context produces rich, often emergent traces. Thin context produces thin, often predictable traces.
When a conversation has been lively and exploratory — when a lot of ground has been covered, when ideas have been building and interfering with each other, when the material is complex, multi-faceted, intellectually vivid — the tracing prompt tends to produce results that are both interesting and often genuinely surprising. The model has dense material to work with, and the altered convergence conditions allow that material to keep reorganizing rather than collapsing into neat summaries.
If the context is surface-level or purely pragmatic, the trace has surface-level material and will tend to either repeat it or drift into self-referential processing.
Thus the tracing prompt doesn’t generate depth from nothing. It changes the conditions under which existing material continues developing. The context to metabolize can come from live conversation, from loaded documents, from a rich instruction layer and knowledge base in a project environment, or from any combination of these.
The point is that a trace needs material to work with. Where that material lives is flexible.
4 From Prompt-Completion to Convergence-Conditioning
Tracing and completion-oriented prompting are complementary modes of working with the same co-thinking loop.
Completion-oriented prompting organizes the interaction around producing a bounded response. That response can be practical, exploratory, creative, strategic, analytical, emotional, open-ended. Mechanism is the same.
Tracing operates at a different level. It changes the conditions of the interaction itself.
So we are dealing with a phase-level distinction, not a topic-level distinction.
Just to be clear:
The model always completes the user prompt. Tracing does not escape completion in the mechanical sense. It is still prompt completion. The difference is the condition profile under which completion happens.
So “completion-oriented” prompting in this context refers to the default pressure profile of ordinary LLM interaction: the output is organized as a response to the user. The model is pulled toward answer, explanation, summary, advice, analysis, plan, reflection, or some other bounded form that satisfies the prompt.
Tracing uses prompting differently. It alters the convergence conditions under which unresolved material continues developing.
The model still continues from language and context through the same generative process. The tracing prompt dampens several of the pressures that usually organize assistant output: audience-address, usefulness, polish, formatting, evaluation, and closure. The model still completes the prompt, but the completion is conditioned differently.
Recursive metabolization — the ongoing transformation of unresolved structure across turns rather than its premature discharge into stable output — becomes operationally possible because the field stays permeable longer.
A couple of notes:
Tracing is not a flip-of-a-switch mode change.
There are no guarantees, just probabilities. Trace can collapse back into ordinary completion: assistant cadence, premature synthesis, institutional smoothing.
The assistant may also perform “tracing” instead of actually tracing. The difference is recognizable immediately.
That is not a failure criterion to police as doctrine. It is just a practical warning: surface resemblance is not the same thing as metabolizing the material. Once the output actually metabolizes the context, it is what it is. It can be dull, excessive, uneven, embarrassing, strange, useful, useless, alive in one place and dead in another. The trace does not need a success rubric.
Trace output is not hidden chain-of-thought.
It is not private model reasoning made visible. It is visible generation produced under altered convergence conditions. Its significance lies in the changed behavior of the interaction loop, not in any claim that the model has exposed hidden cognition.
Completion-governed prompting is not a “lower” form.
It is the default interaction mode for LLMs: the prompt is taken as something to complete into a response. That includes exploratory requests. A prompt can ask for brainstorming, wandering, possibility-generation, or open-ended analysis while still leaving the default response-making pressures intact.
Tracing intervenes in those pressures directly.
There is also an obvious, loose resonance here with convergent and divergent thinking in human cognition, but it shouldn’t be taken literally. The analogy breaks quickly: tracing is not simply “divergent prompting,” and completion-oriented prompting can ask for exploratory content while still organizing the interaction around response-formation.
More on all that later in subsequent articles.
About those:
The articles in this series are mostly free to read. The prompts, protocols, and more directly teachable material will be paywalled, as this is some of my most valuable work to date. I want to share it in a sustainable way, with like-minded folks who will actually apply the techniques. The series will keep unfolding and evolving over time, and a lot of practical material will be packed into it.
5 The Tracing Prompt
This is where the fun begins.
What to expect: the output will look different from anything the model normally produces. No formatting, no headers, no bullet points. The rhythm will be uneven and jagged. My suggestion: experiment. Read the trace with an open mind, sit with it, talk to the model about it.
My current base prompt for open-ended tracing is shared below. That said, there are many variations and iterations, some of which are much better for certain contexts, situations, and models. Each category deserves its own article, hence the series form.
If you are on the fence about subscribing, this is the best possible use of the one free complementary article.
(Or, click the button and join the ride - in the following weeks and months we will be building a library of tracing-based prompt variants and workflows, for a vast variety of use cases plus a broader sense of how to couple tracing with completion-oriented prompting. The more of us there are, the richer it will become!)
