This exercise introduces participants to one another and to the project's exploration of the social and material contexts in which professionals interact with LLMs. Through spontaneous and deliberately imperfect sketches, participants capture their intuitive and metaphorical understanding of LLMs.

This exercise helps participants reflect on their existing work, revealing overlooked or underappreciated tasks.
They also explore what other work they could do with the assistance of an LLM, generating material for future exercises.

This exercise prompts participants to critically assess their LLM interactions through self-evaluation and a structured review of their conversation history.
It highlights gaps between perception and actual use to further their understanding of their LLM usage rather than relying on habitual talking points or anecdotes.

Participants analyze past LLM interactions that stood out due to emotional impact, moral hesitation, or significant outcomes.
They rate the model's performance and explore why it performed as it did, fostering awareness about LLM capabilities and professional expectations.

This thought experiment creates artificial constraints on how participants talk about their use of LLMs.
Because they cannot rely on ready-made categories (e.g., 'ChatGPT', 'machine', or 'assistant'), participants need to reconsider their interactions with the models, playing with different metaphors and analogies drawn from their social and professional environments.

This exercise guides participants through a selection process, drawing from their current work practices and the potential implementation of new tasks.
In contrast to standardized benchmarks, which are tailor-made by engineers for LLMs, participants design a trial that revolves around their professional experience.
This personalized benchmark aims to assess whether the models can provide practical value that actually matters in a specific professional situation.

By formulating detailed instructions, participants reflect on how they can effectively guide an LLM toward accomplishing a simple or complex task. They are led to consider the weight of their own prompt on the final result.
By selecting a human or a specialized model, participants explore different avenues of comparison - machine versus machine or human versus machine.

In this exercise, participants enlist the help of a fellow co-inquirer, a good friend, or a colleague.
In addition to anonymizing results and ensuring fairness in the upcoming trial, this exercise allows participants to concretely experience the collaborative nature of LLM performance evaluation.

Participants rank anonymized LLM responses for each task, determine the best-performing model, and critically review the trial's fairness and relevance.
This evaluation encourages participants to reflect on what matters in assessing the situated value of LLMs.

Participants explore a specific prompting technique by practicing it throughout the week, and summarizing their findings in a concise presentation. The goal is to develop a nuanced understanding of how prompt design influences LLM responses.
Knowledge acquired about a prompting technique is shared among all participants through a short presentation, based on observations they kept track of in a personal diary.

Participants document how their use of LLMs has changed since the beginning of the experiment. They write two short essays—one on how their practices have evolved and another on what has remained stable—using concrete examples from their experience.
Through a structured discussion game, participants also analyze their shifting perspectives on LLMs. The goal is to engage in collective reflection in order to challenge assumptions and reinforce insights gained through the experiment.

An exemplary work is not necessarily a prestigious publication but also modest, everyday writings that embody clarity, precision, or impact within a professional context. By selecting both renowned and less conspicuous examples, the goal is to sharpen awareness of what good writing looks like across different formats and purposes.

Participants attempt to recreate a high-quality text using an LLM without providing direct access to the original. This experiment tests the LLM's ability to replicate style, tone, and meaning while challenging participants to refine their prompting strategies.
Through 90-minute work sessions, screen recordings, and a logbook, participants document their process and assess the model's capacity to (re)produce a text it does not have access to.

This exercise contextualizes the chosen exemplary work by examining its professional significance. A radar graph helps assess factors like frequency, expertise required, and error tolerance, while a structured reflection highlights what makes the work exemplary and worth emulating.

Participants analyze the underlying structure of their chosen text, mapping its key influences, professional constraints, and dependencies.
A layered dependency graph helps distinguish which elements are accessible to an LLM and which remain beyond its reach, reinforcing an understanding of AI's limitations.

By identifying challenges that can be overcome through iteration or adapting a prompt, situations where the LLM fails to generate meaningful progress despite repeated attempts, and effortless response generation, participants discover patterns in their experiences and develop a clearer sense of when to persist, pivot, or abandon a strategy.

By mapping progress horizontally and obstacles vertically, the drawing becomes a living diagram of how one navigates challenges, setbacks, and breakthroughs while learning a new skill with the help of an LLM. It encourages an intuitive, embodied reflection on learning —where success isn’t linear, and detours or derailments become part of the story.

Rather than summarizing past exercises, this exercise encourages the creation of a personal booklet that combines selected pages from the vademecum with additional notes, sketches, and digital traces. This curated archive serves as both a tangible memory and a storytelling device —capturing key moments, doubts, and insights that have shaped the participant’s evolving relationship with LLMs.