theme-ai-support-for-reflection

Started out trying to write a paper that responds to this, but as it is currently (11/05/2026) many of the AI/LLM implementations are still speculative with only some first prototypes made, I don’t feel I have substantial material to make this a good paper.

As such I am going to respond to this theme theme-creativity-support-tools-for-reflection reworking my live-coding-a-phd written work, to talk about how liveness, performance and making public make this platform a CST for reflection.

theme outline:

AI models that respond to complex, open-ended, and multimodal inputs at scale bring forth unique opportunities and challenges for supporting creative practitioners’ reflective thinking. Concerns persist regarding the potential negative cognitive effects of AI use. For instance, there is a cognitive process mismatch between creators and LLMs: LLMs typically operate through direct, goal-directed reasoning whereas creative practice is inherently non-linear and characterised by loosely defined intentions.

• How can creators contend with the cognitive mismatch between LLMs and their own thinking?
• Can structure be implemented in LLMs to scaffold creative reflection without impacting characteristics of the creative user experience?
• Will creators retain the opportunity to reflect on surprises and follow tangents within an AI’s constraints?
• Will creators feelings of agency and creative intent be preserved when working with reflective AI scaffolds?

drafting:

We particularly encourage submissions of first-person reflective pieces on people’s own computer arts practice, especially with a brief meta-analysis (1 paragraph) of how reflection occurred in their making.* –> where does this go?

What does it look like to build a personal research infrastructure that uses AI, but is designed around how you actually think rather than around how AI wants you to think?

Building around the thinking: a personal AI research infrastructure designed for non-linear creative practice

setting out to research generative AI in learning and education, I was not expecting the site of my growing research to become in quite a meta sense - the core site of my research. But from the start rather unintentionally the foundations of a wider personal AI research infrastructure was seeded.

At it’s base, my notes and thinking are structured around a zettelkasten-inspired approach, which naturally allows for the slow formulation and clustering of ideas. I began to wonder whether from this emergent cluster there could be ways to implement LLM tools to surface and work with this material alongside my thinking, but always keeping them within a seperate domain. This is the central tension I am trying to understand and build toward: how can I work with LLM’s on the growing dataset that is my research, while working against some of the understood behaviours and functionalities of this technology - optimisation, productivity and efficiency and swift evaluation. How can I instead create a space where thoughts can be investigated and played with, where AI stimulates and aids my process rather that taking the reins from me by too swiftly resolving what should remain open

As my research began to grow, naturally I started to see and understand that this was a perfect dataset for LLM tools that might aid me in my research and that I should built on top of this.

the zettelkasten instinct

• started from an arbitury place of trying to organise my PhD in a good way that would benefit me over the span of the project
• discovered zettelkasten as an administrative solution that unexpectedly opened into something conceptually richer – ideas growing and clustering through association rather than hierarchy
• this led to a tangential imaginative question: what if you could live code a PhD?
• that framing connected to play, to speculative thinking, to process over outcome – it wasn’t a methodology(RQ3-playful-methodology-as-way-to-speculate explores playfulness so this is a thing that has been in my mind since the start), it was an imaginary that then became a system

The system as it exists

• two layers working in parallel: Obsidian vault as the protected space of actual thinking – chaotic, associative, non-linear – and PhD-Live as the public-facing surface that makes that thinking more legible without flattening it

• they are distinct but connected – not the same thing presented differently

• noticed something interesting in practice: often reading through posts on the PhD-Live website rather than inside Obsidian – the two environments produce different relationships to the same material. When I am ‘outside’ of sessions I would read over and think about ideas on the website
  -it helps me see areas that are ‘lacking’ and need to be developed, perhaps that awareness of legibility required for a website?

  • the live layer logic of tracking recent sessions and thought allows me to more readily recoginse progress and thread as they develop and mature

• the public layer makes the thinking readable in a way that the vault doesn’t, but the vault is where the thinking actually lives

• This two-layer structure wasn’t planned – it emerged from use, and only became visible as a design decision retrospectively
  The design question: where AI enters and where it can’t

• strong instinct from early on that the LLM implementation must not happen directly inside the Obsidian vault – they can be connected, but the preservation of how I think is important to keep intact

• what I’m imagining is a third layer: AI that can surface patterns and alternate understandings from the base knowledge, similar to how Obsidian’s graph view surfaces growing nodes – emergent connection rather than directed reasoning

• the critical constraint: what the AI surfaces must not state my thoughts for me – it can reflect, cluster, make visible, but it cannot resolve or speak

• this is the design tension I’m working inside: how do you build an AI scaffold that participates in associative thinking without collapsing it into something more linear and goal-directed than it should be?

what did I actually build, and what did building it reveal?

context:
working with obsidian as the core container, I began working with this as a way to track my thoughts, inspired by Zettalkasten, it was proving a useful tool in the very early stages of starting to demonstrate to me the ways in which my seemingly disparate thoughts were beginning to connect and become coherent directions. In this process I was simultaneously reflecting on a thread that is important to my research, which is a playful and imaginative approach that was informing my methodological approach. I was also interested in ideas of performance from my practice, I had this thought – ‘what could you could live code a phd?’. This was to explore a methodical approach to conducting my research, but also a direct question and hypothesis that Ihad in regards to the growing influence of generative AI on education. If the assessed outcomes within learning can all be produced easily with generative AI tools, does it everything become centered on the process? This system came to embody and explore both of these possibilities.

build / how it works / capturing liveness:
early on in my research alongside learning more about zettelkasten, the idea of capturing research as it happens lead me to the concept and practice of digital gardens. It so happens that obsidian is a tool that many use for keeping digital gardens, so I found quickly an implentaion for making public my obsidian notes to a garden, that I used as the main scaffold for doing this. From there there question was, how can I display my progess in a live way. I designed a system into this framework that would, as naturally as possible capture my process and progess and publish this live.

(excerpt from previous paper describing the site)
To capture and scaffold my work I created a simple typology of differentiators that capture the varying scope of activity that I might be doing in my research. The system distinguishes between
several note types within Obsidian that translate into different behaviours on the website: daily notes as chronological working logs, sessions as focused periods of activity embedded within or alongside these dailies, posts as later-stage syntheses for public reading, and general notes as atomic, conceptfocused
entries aligned with a Zettelkasten-inspired practice of building networks of ideas. Together, these layers create a typology that separates raw process, sustained work, and reflective output
while keeping them computationally linked.

Alongside this, I have implemented a ‘live layer’, that surfaces what I am doing in real time, this operates as a lightweight activity-tracking system embedded directly within my everyday notetaking practice. Short, timestamped entries, inline session markers, and milestone annotations are parsed automatically during site builds and aggregated into a public-facing dashboard that present what is happening now, what has recently occurred, and what is planned. Rather than relying on manual logging or separate tools, liveness emerges from ordinary writing gestures inside notes. This approach treats the documentation itself as a performative research act, collapsing the boundary between working, recording, and publishing. The system is designed to integrate as seamlessly as possible into my existing workflow, allowing the liveness of my work to be captured as it emerges and develops. This approach aims to avoid the performative pressure of presenting work differently for an audience, which risks turning the liveness the system seeks to register into an artifice. The public facing website is minimal in design, feeling like a notebook - to focus on the forming ideas as the main event, the emerging thoughts taking centre stage.

what emerges from this:
once built it started to become clear how the system having this public facing garde was startign to impact my research process. obsidian became like my sketchbook so to speak, the place for chaotic and rough notes and place to start capturing the connections, a the public facing side started to make the core ideas come through more clearly, it gave me the opportunity to read my notes through a different lens, less close to it with a detachment as if I was a member of public which also allows me to see where things are not very legible need further developing or expression, I wasnt expecting that at all. I noticed that I would read over my note more often on the public site that with obsidian and it was interesting that even with this simple platform build, the two environments (private obsidian vault and public website) would allow me different relationships with the same material. I was not planning this at all and it emerged through use.

back to the the question of AI:
so the the building of the site was in direct part a response to the way that things are changing in academic institutes in response to generative AI. But inthe process of the platforms purpose and abilties to shape my process, I was thinking also about AI, there is this tension of LLM’s impact on academic work and I started to think, what if I could make that impact live and and emergent too through this platform, what if you could see its impact and what if AI had the ability to be another strand of ‘reflection’ that could happen within the the site. Importantly however, working within my obsidian, and seeing these nodes and connections grow organically, its become clear to me the value in this and that bringing in an LLM into this space could (in my view compromise it). but the value in the notes clusters and emerging thoughts in my obsidian vault is not the fact they are perfect and polished thoughts, but the tracking, collection and connecting of ideas and cohesion over time, a lot of the time through a reflective practice of revisiting the notes (along side my making practice and contextual resarch). It seems to me that the general use case of LLMs is in a solutionist way (mention Karpathy’s LLM wiki?), but it within this environment that kind of functionality is not welcome. However alongside this emerging, ‘digital garden’ space, it feels potentially useful to have the LLM also ties together emergent threads and reflect patterns that are emerging - purely based off of the work and knowledege that I am creating myself. In this context it’s role is defined in a supportive role. The core question is how to scaffold this behavior in a way that is truly supportive of this and doesn’t collapse into something more linear and goal-directed than it should be.

The reflection: grappling with the two RiCE questions

• “how can creators contend with the cognitive mismatch between LLMs and their own thinking?” – my working answer is: by designing the system architecture so the mismatch is preserved rather than resolved. the vault stays chaotic. the AI operates on the legible surface, not the source.
• “will creators’ feelings of agency and creative intent be preserved when working with reflective AI scaffolds?” – the instinct that the AI must not state my thoughts for me is itself an answer to this. agency is preserved by being deliberate about what the AI can reach and what it can’t.
• neither of these is a finished answer – the AI component isn’t built yet, and that incompleteness is honest. I’m grappling with whether and how to bring AI into the system precisely because of these questions.
• the not-yet-built-ness is part of the argument, not a weakness