Critical Code Studies Workbench

Comments

• jshrager

  4:18AM

  Oh, very cool. This inspired me to this: https://claude.ai/share/083376a6-3648-40dd-a097-d8841259899d
• markcmarino

  7:38AM

  @jshrager And what do you think of its reading?

  I am glad that LLMs can reply to these sorts of prompts, personally. @jeremydouglass and I were just remarking on the advantages of these models having trained on texts of Critical Code Studies, though of course not always with permission.

  Thinking back to Week 1, I wonder if we might get a little more out of choosing a particular critical lens for it to apply.

  Maybe if you have it suggest some approaches from various schools of critical theory and the choose one for it to approach.

  But that does make this interpretation business into a bit of a game when I think it usually begins with a question or even a hunch. A critical inquiry. A focus of reflection that gets crystalized when it hits an object of study.

  For example, there are lots of critical approaches that reflect on identity (like the psychoanalytic approaches: Lacan, Irigaray -- not to forget a Rogerian approach!) or post-humanism (Hayles) or cyborg theories (Haraway). As you know from our book, I have chatbots through the lens of gender performance (Butler). I know you also like theories from cognitive and neuroscience, so maybe try one of those.

  In other words, the word hermeneutic might now be sufficient without pointing it to a particular hermeneutic approach.

  If you ask the LLM, what parts of this code could be used in a critical code studies reading of this basic ELIZA in terms of (then choose one of those theories or theorists to fill in this blank) as I reflect on (fill in this blank with some philosophical or critical question), what does that yield?
• jshrager

  8:04AM

  @markcmarino Re: What I thought of Claude reading my code ...

  This was sort of interesting:

  "But here's the hermeneutic violence: the program assumes that merely swapping pronouns creates meaningful reflection. If you say "I FEEL you are wrong", it becomes "I FEEL I am wrong" - a grammatical transformation that may create psychological non-sense."

  I love the term "hermeneutic violence"! :-) However, I feel that it's wrong that "I feel I am wrong" is a nonsensical response to "I feel you are wrong"; The speaker might have convinced the listener, or the listener might be musing over whether they actually feel they are wrong.
• davidmberry

  11:06AM edited 11:38AM

  In creating the critical code studies workbench, I thought it would be useful to distill what a critical code studies approach/method would be into a "skill" for the LLM. This means I gave an LLM a collection of CCS texts to read and asked it to convert their contents into a knowledge skill (rather like knowledge elicitation for an Expert System) for the AI so that it could apply CCS in practice. This hones its ability to focus on the task within an expert domain, rather than just ask a vanilla LLM to apply a (random) CCS method. This dramatically changes the ability of an LLM to understand what you want, and suggests a good way to do it.

  In the end this Skill file proved too large (particularly for Ollama) for I had to compact it rather dramatically into this version (note this is also designed for progressive loading of the right skillset for the right mode in the workbench).

  But I think this is still a useful distillation of the critical code studies methodology and you can paste it into an LLM to "teach" it how to do a CCS reading to preload it, as it were.

  ---
  name: critical-code-studies
  description: Comprehensive methodological framework for Critical Code Studies (CCS), integrating materialist-phenomenological method, hermeneutic-rhetorical approach, and centrifugal close reading from 10 PRINT. Use when analysing code as cultural text, conducting close readings of software, or applying critical theory to computational systems.
  version: 1.0.0
  date: 2026-01-20
  author: David M. Berry
  ---
  
  # Critical Code Studies Methodology
  
  ## Overview
  
  This skill provides a comprehensive methodological framework for Critical Code Studies (CCS), integrating four complementary approaches:
  
  1. **Materialist-phenomenological method**: Code as crystallisation of social formations, examined through multi-dimensional analysis (literature, mechanism, spatial form, repository), tests of strength, political economy, and phenomenology of computation.
  
  2. **Hermeneutic-rhetorical approach**: Code as social text with extrafunctional significance, analysed through critical hermeneutics, close reading, rhetorical analysis, and interpretation of meaning beyond functionality.
  
  3. **Centrifugal close reading (10 PRINT)**: Spiral outward from single line or minimal program to explore seemingly disparate aspects of culture, treating code through variorum approach across different printed variants and platforms.
  
  4. **Critical lenses (race, gender, postcolonial)**: Code examined through feminist theory, critical race theory, and postcolonial criticism, attending to how computational systems encode and reproduce social hierarchies, construct gendered and racialised subject positions, and extend or resist colonial logics.
  
  These approaches share commitment to understanding code as embedded in power relations, requiring critical engagement with technical operations and social contexts whilst recognising code's extra-functional significance and cultural circulation.
  

  The full Critical Code Studies LLM Skill is here

  The idea that you can take a collection of books and papers and convert it into a "skill" for an AI is super interesting. I think it is also suggestive of how CCS can operationalise its approach and method, particularly for teaching and research, where it can be difficult to understand how it is meant to function.

  Of course, this skill is also extendable so that multiple approaches to CCS can be "taught" to it via an LLM (or hand edited by a human). I can imagine a number of improvements that could be made to this 1.0 version of the CCS skill – but this would also require compaction as current LLMs have a limited context window (250k for Clause, 1 million for Gemini) and the bigger this file gets the more tokens are used in parsing it by the LLM – hence the contraction I referred to earlier for the CCS workbench (as Ollama has a much smaller context window, 128K tokens).
• anthony_hay

  11:15AM

  Hi @jshrager Is Claude's interpretation correct? Wouldn't the program respond to "I FEEL you are wrong" with only one of these?

  "DOES THAT TROUBLE YOU?"
  "TELL ME MORE ABOUT SUCH FEELINGS."
  "DO YOU OFTEN FEEL I AM WRONG"
  "DO YOU ENJOY FEELING I AM WRONG"
• jshrager

  3:30PM

  @anthony_hay I don't know. I'd have to run the program to find out exactly, but I don't think that this matters much. If you and I were talking about this sort of thing, I wouldn't expect you to be able to generate exactly correct examples to make this sort of point. In fact, it's this precise sort of flexibility that makes LLMs so useful, and at the same time so frustrating. Often we want some flexibility -- creativity, if you like -- say in envisioning the UI of something you're building, but when it comes time to implement, we want to drill down and get the details right -- a process sometimes called "cognitive zoom", but these have no capability to engage in cognitive zoom, so they treat programming the same way they treat poetry. As a result, possibly 2/3rds of the interactions I have had over several years of AI-assisted coding involves getting the system to find stupid mistakes that no real person would make, like not balancing their divs. (To be fair, a real person might make such a mistake, possibly more often than an LLM if left to their own devices. But "real" engineers know that this sort of thing is important and so we have specific machinery building into our IDEs that do this sort of thing for us. This is the core of the problem; LLMs depend on fancy statistical processes for EVERYTHING. Now, again, to be fair, the AI folks realize this, but unfortunately, aside from paren balancing and a couple of other things, like arithmetic, it has been previously demonstrated (actually not, but that's a separate thread) ... it is believed, let's say, that we can't build such "expert systems" for complex tasks. So what LLMs have basically done is to reduce AI to a previously unsolved problem.)