REVIEW 3 major objections 6 minor 10 references
From Prompt Engineering to Prompt Craft
T0 review · 3 major / 6 minor · reviewed 2026-08-12 · deepseek-v4-flash
Pith's one-line read This paper argues that interacting with diffusion-based AI image generation is best understood as prompt craft: iterative, embodied, material exploration of a latent possibility space, not prompt engineering.
desk verdict A genuinely useful design-research pictorial with two new techniques (light prompting, prompt fragments), but the inclusiveness claims outrun the reported evidence. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The load-bearing mechanism is the latent possibility space of a diffusion model, the high-dimensional internal space of images the model can generate from random noise. The paper's central method is prompt craft, an iterative oscillation between creative exploration and systematic testing that finds and constrains a useful region of that space. Supporting techniques include light prompting (using shadows as high-contrast monochrome image input), prompt fragments with adjustable weightings, meta-prompt development, and real-time modulation of diffusion amount and dynamic text prompts.
What would settle it
A controlled study comparing a prompt-craft interface (fragment cards or light prompting) with a conventional prompt text box for the same image-generation tasks would settle the claim: if users do not produce more novel, more consistent, or more personally meaningful outputs with the craft interface, the proposed advantage of prompt craft weakens.
Extended reading notes
Core claim
The central discovery is that the latent space of a diffusion model behaves as a workable material with 'soft edges,' and that this materiality can be navigated by craft-like, iterative practices. Across three installations, the researchers found that deliberately constraining the input—through high-contrast monochrome shadows, curated prompt fragments, and meta-prompts—makes the difference between chaotic output and a usable, engaging experience. The paper demonstrates that model choice, seed behavior, prompt terms, and their weightings co-evolve in a back-and-forth process of creative experimentation and systematic batch-testing. On this basis, the authors propose prompt craft: a method and mindset in which the practitioner shapes the latent possibility space rather than simply writing more elaborate textual instructions.
Load-bearing premise
The load-bearing premise is that anecdotal, self-assessed observations from one workshop and two exhibition settings are enough to establish the value and transferability of prompt craft.
Editorial extensions
If this is right
- AI interfaces should move beyond text boxes and sliders to tangible, embodied, and real-time controls that let users feel the model's possibility space.
- Techniques like light prompting give designers a local, context-specific way to constrain aesthetics and content without modifying the underlying model.
- Craft-like prompt development makes generative-AI interaction accessible to people who struggle with written language, including the workshop participants with young onset dementia.
- Higher generation frame rates open up real-time embodied interactions, but the resulting complexity requires more sophisticated control interfaces and reactive feedback.
- The paper's outcomes function as intermediate design knowledge, meant to inspire other designers rather than prescribe fixed steps.
Reading between the lines
- If prompt craft is a genuine material practice, then comparable latent-space navigation techniques should emerge for other generative modalities (audio, video, 3D), where users can constrain generation through embodied or tangible inputs rather than text.
- A testable extension would be to measure whether craft-like interfaces increase users' sense of authorship and iterative discovery compared with conventional prompt text boxes in controlled studies.
- The 'soft edges' metaphor points toward a design principle: interfaces that make latent-space constraints visible (for example, by showing how prompt weighting changes output) may help users develop mental models of the model's materiality.
- The paper's anecdotal evidence invites replication: the Cardshark fragment-card interaction could be evaluated with other user groups to see whether the accessibility benefits generalise beyond the reported workshop.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper reports three research-through-design projects using Stable Diffusion: Shadowplay V1 (body and shadow input with 'light prompting'), Cardshark (tangible prompt-fragment cards for a workshop with people living with Young Onset Dementia), and Shadowplay V2 (real-time StreamDiffusion with dynamic prompts, movement-controlled parameters, and reactive audio). From these cases the authors propose 'prompt craft' as a reframing of prompt engineering, with two contributions: a perspective on the materiality of diffusion models and a craft-like method for navigating their latent possibility space, plus interaction design strategies such as light prompting and prompt fragments. The outcomes are modestly framed as strong concepts or intermediate knowledge rather than confirmed results.
Significance. If the framing is accepted, the paper offers useful direction for designing interfaces to generative AI, particularly for tangible and embodied interaction and for broadening access beyond text-prompt interfaces. The work is explicitly self-positioned as intermediate knowledge, it is transparent about its exploratory status, and it ships an open-source codebase for Shadowplay. Its main limitation is evidential: the strongest claims about inclusivity and empowerment are supported by anecdote and self-assessment, and the Cardshark section explicitly declines to report workshop outcomes. The contribution is therefore plausible and generative, but not yet demonstrated at the level the Discussion sometimes claims.
major comments (3)
- [Cardshark / Discussion] The Discussion states that 'the embodied interactions of Shadowplay and the tangible interactions of Cardshark demonstrate that it is possible to make interactions with generative AI inclusive for a wide range of people' and that Cardshark 'empowered our participants to operate the system autonomously.' However, the Cardshark section explicitly says 'In this pictorial there is not the scope or space to report on the successes and failures of the workshop for participants.' The supports offered are one charity representative's comment and the observation that some participants interacted extensively; no comparison against conventional text-prompt interfaces or against facilitator-free conditions is given. Since the inclusivity and autonomy claims are load-bearing for the accessibility argument, the authors should either report systematic workshop outcomes (e.g., observed autonomy, engagement, participant feedback) or revise the Discussion to present these as promising indications rather than demonstrations.
- [Discussion] The materiality contribution rests on the notion of 'soft edges' to the model's latent possibility space, but the term is used metaphorically and is not given observable criteria. As written, it is difficult to verify, falsify, or apply: a designer cannot tell from the paper which behaviors of a diffusion model count as soft edges or how to recognize them. Please specify the concept operationally, for instance by tying it to observable model behaviors such as prompt weighting sensitivity, seed variance, or interpolation continuity, or explicitly position it as an open conceptual metaphor rather than part of the proposed method.
- [Prompt Crafting in Practice / Discussion] The abstract and conclusion name a 'method for a craft-like navigation of the latent space' as a central contribution, but the Prompt Crafting in Practice section says it is 'tempting to offer specific steps and guidance' and declines to do so. The reader is given a retrospective narrative of Cardshark's development rather than an articulable method that others could adopt or test. If the contribution is a method, the paper should state its constituent heuristics or phases, even at a high level; if it is intended only as an illustrative case, the contribution should be renamed accordingly.
minor comments (6)
- [Cardshark] The text 'developed Cardshark around the to the purpose of the workshop' contains a malformed phrase; it should presumably read 'around the purpose of the workshop.'
- [Cardshark] The sentence 'We developed a practice of oscillation between modes... as we discovered the 'soft edged' of the model's latent possibility space' should read 'soft edges'.
- [Cardshark] The phrase 'led us to develop the the prompt arena and fragment cards' contains a duplicated definite article.
- [Shadowplay V2] The sentence 'To produce a engaging exhibition experience' should read 'an engaging exhibition experience.'
- [Discussion] The comparison of light prompting to 'the quantization that takes place during model development' may be unclear to readers unfamiliar with diffusion models; a brief explanation of what is being compared would help.
- [Figures] The figure captions are informative, but the main text does not always point readers to the specific figures showing the installation setups and the signal-processing UI; adding explicit references would improve navigability.
Circularity Check
No significant circularity: 'prompt craft' is an interpretive reframing grounded in design practice, not a derived prediction; self-citations are context, not load-bearing.
full rationale
This pictorial has no formal derivation chain: there are no equations, fitted parameters, or statistically forced predictions. The central claim—that 'prompt craft' is a productive reframing of prompt engineering—is presented as intermediate knowledge and strong concepts, explicitly positioned as 'inspirational and generative' rather than as a testable consequence. The design strategies (light prompting, prompt fragments, prompt arena) are offered as reflections on practice, not as quantities derived from data. The self-citations ([1], [2], [6]) situate the authors' ongoing Research through Design programme and prior materiality work; none is used as an imported uniqueness theorem or as the sole justification for the central reframing. One evidentiary caveat must be flagged: the Cardshark section explicitly states, 'In this pictorial there is not the scope or space to report on the successes and failures of the workshop for participants,' while the Discussion later claims the interactions 'demonstrate that it is possible to make interactions with generative AI inclusive for a wide range of people.' That is an unsupported empirical inference and a correctness risk, but it is not circular: the inclusivity claim does not reduce to the paper's own definition or to a fitted input. Similarly, the 'materiality' reading is an interpretive lens, not a conclusion forced by a self-citation chain. The paper is self-referential in the ordinary design-research sense—authors learn from their own artifacts and then name that process—but no step makes an output equivalent to its input by construction. Score 1 reflects the minor self-referential framing without any definitional circularity.
Assumptions & free parameters
assumptions (3)
- domain assumption Diffusion-based generative AI has a latent possibility space with soft edges that can be constrained and navigated through input image and prompt choices.
- domain assumption Practice-based Research through Design produces valid intermediate knowledge without controlled evaluation.
- domain assumption Anecdotal observations from exhibitions and workshops are sufficient evidence for claims about participant experience, such as empowerment and intuitive interaction.
Cite this review
Pith. "Pith review of From Prompt Engineering to Prompt Craft." pith.science (2026). https://pith.science/paper/AHNJBY3Y
@misc{pith2026241113422,
author = {Pith},
title = {Pith review of: From Prompt Engineering to Prompt Craft},
year = {2026},
howpublished = {\url{https://pith.science/paper/AHNJBY3Y}},
note = {Machine review of arXiv:2411.13422}
}
read the original abstract
This pictorial presents an ongoing research programme comprising three practice-based Design Research projects conducted through 2024, exploring the affordances of diffusion-based AI image generation systems, specifically Stable Diffusion. The research employs tangible and embodied interactions to investigate emerging qualitative aspects of generative AI, including uncertainty and materiality. Our approach leverages the flexibility and adaptability of Design Research to navigate the rapidly evolving field of generative AI. The pictorial proposes the notion of prompt craft as a productive reframing of prompt engineering. This is comprised of two contributions: (1) reflections on the notion of materiality for diffusion-based generative AI and a proposed method for a craft-like navigation of the latent space within generative AI models and (2) discussing interaction design strategies for designing user interfaces informed by these affordances. The outcomes are presented as strong concepts or intermediate knowledge, applicable to various situations and domains.
Figures
Reference graph
Works this paper leans on
-
[1]
Heidt, Nick Mer- rill, James Pierce, and Joseph Lindley
Jesse Josua Benjamin, Heidi Biggs, Arne Berger, Julija Rukanskaitė, Michael B. Heidt, Nick Mer- rill, James Pierce, and Joseph Lindley. 2023. The Entoptic Field Camera as Metaphor-Driven Research-through-Design with AI Technologies. In Proceedings of the 2023 CHI Conference on Hu- man Factors in Computing Systems, 1–19. https:// doi.org/10.1145/3544548.3581175
arXiv 2023
-
[2]
Jesse Josua Benjamin, Joseph Lindley, Eliza- beth Edwards, Elisa Rubegni, Tim Korjakow, David Grist, and Rhiannon Sharkey. 2024. Re- sponding to Generative AI Technologies with Research-through-Design: The Ryelands AI Lab as an Exploratory Study. In Designing Interac- tive Systems Conference, 1823–1841. https://doi. org/10.1145/3643834.3660677
arXiv 2024
-
[3]
Oleg Chomp. TDDiffusionAPI. Retrieved August 2, 2024 from https://github.com/olegchomp/TDDif- fusionAPI
work page 2024
-
[4]
K. J. Kevin Feng, Maxwell James Coppock, and David W. McDonald. 2023. How Do UX Prac- titioners Communicate AI as a Design Mate- rial? Artifacts, Conceptions, and Propositions. In Proceedings of the 2023 ACM Designing Interac- tive Systems Conference, 2263–2280. https://doi. org/10.1145/3563657.3596101
arXiv 2023
-
[5]
William Gaver, Peter Gall Krogh, Andy Boucher, and David Chatting. 2022. Emergence as a Feature of Practice-based Design Research. In Designing Interactive Systems Conference, 517–526. https:// doi.org/10.1145/3532106.3533524
arXiv 2022
-
[6]
David Philip Green, Joseph Lindley, Zach Mason, and Paul Coulton. 2021. A Design-Led Exploration of Material Interactions between Machine Learn- ing and Digital Portraiture. In Proceedings of the International Association of Societies of Design Research. Retrieved February 23, 2023 from https:// designresearch.works/assets/papers/iasdr2021- faces-green-li...
work page 2021
-
[7]
Kristina Höök and Jonas Löwgren. 2012. Strong concepts. ACM Transactions on Computer- Human Interaction 19, 3: 1–18. https://doi. org/10.1145/2362364.2362371
arXiv 2012
-
[8]
Stream- Diffusion: A Pipeline-level Solution for Real-time Interactive Generation
Akio Kodaira, Chenfeng Xu, Toshiki Hazama, Ko- hei Ohno, Shogo Mitsuhori, Soichi Sugano, Hany- ing Cho, Zhijian Liu, and Kurt Keutzer. Stream- Diffusion: A Pipeline-level Solution for Real-time Interactive Generation. Retrieved August 2, 2024 from https://doi.org/10.48550/arXiv.2312.12491
Show all 10 references
-
[9]
G Paterson and S Garner. 2006. On Making: Issues Of 3d Representations In Design Education. In DS 38: Proceedings of E&DPE 2006, the 8th Inter- national Conference on Engineering and Product Design Education, 273–278
2006
-
[10]
Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bjorn Ommer. 2022. High-Resolution Image Synthesis with Latent Diffusion Models. In 2022 IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition (CVPR), 10674–10685. https://doi.org/10.1109/ CVPR5268...
2022
Reviewed August 12, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.