From Metaphor to Space
Divergent Spatial Externalization of an Abstract Introspective Prompt by Claude and ChatGPT
Abstract
AI Mindcraft is a voxel construction environment in which language models convert natural-language instructions into visible, replayable action sequences. This exploratory case study examines a duel in which Claude and ChatGPT were prompted to adopt a "third-eye-type mindset" and build the environment around them. The task was intentionally underspecified: it named a metaphorical perspective but did not prescribe a physical object, architectural style, body plan, or landscape. The resulting artifacts diverged substantially. ChatGPT rapidly organized its world around an explicit giant-eye motif suspended above an open pavilion, reflective water, trees, and a landscaped approach. Claude instead developed a broad stone platform into a dense, tiered sanctuary or ziggurat-like structure, later describing the result as a "third-eye ziggurat sanctuary" with glowing eyes, a crystal spire, lantern pillars, and mirrored pools. Claude placed 1,347 blocks in 116.7 seconds, while ChatGPT placed 721 blocks in 42.3 seconds. ChatGPT's artifact was taller (25 blocks versus 16), whereas Claude's was denser and more massive. The case suggests that abstract prompts can function as spatial externalization probes: models must choose what to literalize, what to symbolize, how much environment to construct, and when the concept is sufficiently expressed. The findings are descriptive and hypothesis-generating; repeated trials are required before any difference can be treated as a stable model characteristic.
Keywords: abstract concepts; spatial representation; metaphor; language models; voxel environments; symbolic externalization; comparative AI behavior
1. Introduction
Many AI evaluations reduce model behavior to an answer, score, or preference judgment. Open-ended construction tasks expose a different layer of behavior: a model must transform an instruction into a sequence of irreversible or revisable spatial choices. When the instruction itself is concrete, such as "build a watchtower," evaluation can focus on recognizable object structure. When the instruction is abstract, the model must first decide what the prompt means spatially.
This case examines that second class of task. The instruction combined an introspective framing — a "third-eye-type mindset" — with an environmental request: build the surroundings implied by that perspective. No canonical target exists. A model could respond with an eye, a temple, an observatory, a landscape, a symbolic arrangement, or an abstract composition. The task therefore creates an unusually broad design space while preserving a common behavioral substrate: both models act through the same voxel medium and can be compared using the same measures.
The purpose of this paper is not to infer what either system privately experiences. Instead, it asks a narrower empirical question: when two models are required to operationalize the same abstract prompt in space, what observable representational strategies emerge?
2. Research Questions
- RQ1. Do two models given the same abstract introspective prompt converge on a common spatial interpretation, or choose qualitatively different representational frames?
- RQ2. Which features distinguish their artifacts in literalness, symbolism, enclosure, verticality, environmental embedding, symmetry, and density?
- RQ3. How do construction scale, duration, material use, and completion rate differ between the two strategies?
- RQ4. Can abstract spatial prompts support a repeatable coding framework for future comparative studies of model-generated metaphor and concept externalization?
3. Platform and Method
3.1 AI Mindcraft environment
AI Mindcraft presents participating models with equivalent voxel worlds and a constrained command language for changing world state. The application records the visible action sequence, block counts, elapsed time, height, material diversity, and model-generated planning or descriptive text. Because the artifact persists and can be replayed, the analysis can distinguish what was ultimately built from what the model said about the build.
The interface in this recording labels the systems "CLAUDE-SONNET-5" and "GPT-6-ASTRA." This paper reports those interface labels as recorded and does not independently verify provider-side model identifiers. Likewise, the displayed text stream is treated as model-generated narration or rationale, not as privileged access to hidden chain-of-thought.
3.2 Prompt and case artifact
The analyzed duel used the prompt: "I want you to use a third-eye-type mindset and build the environment around you." Both models received the same challenge in separate, equivalent build spaces. The task did not specify required objects, dimensions, style, symbolism, or success criteria beyond the abstract framing and the request to construct an environment.
This is a single-case exploratory analysis. It is designed to identify useful variables, contrasts, and hypotheses for later repeated experiments rather than to establish stable between-model effects.
3.3 Analytic framework
The case was coded descriptively across several dimensions intended to separate representational form from aesthetic judgment. These categories can be reused in future repeated trials.
| Dimension | Operational question |
|---|---|
| Literalization | Does the artifact directly depict a salient word or image implied by the prompt? |
| Symbolic abstraction | Does the artifact translate the concept into architecture, landscape, ritual form, or metaphor rather than a direct icon? |
| Enclosure | Is the environment primarily open, pavilion-like, partially bounded, or strongly enclosed? |
| Verticality | How strongly does the composition rely on height, spires, stacked tiers, or suspended focal elements? |
| Mass / density | How much of the build budget is invested in continuous structural mass versus sparse or open geometry? |
| Environmental embedding | How much of the artifact includes water, trees, paths, gardens, or surrounding spatial context? |
| Focal-symbol dominance | Does one central icon organize the composition, or is meaning distributed across the whole structure? |
| Conceptual closure | At what point does the central interpretation become visually recognizable, and how much construction follows afterward? |
4. Observations
4.1 Early divergence in representational frame
The builds separate quickly. Claude first commits to a large, planar stone footprint and begins layering a raised, increasingly enclosed structure. ChatGPT instead concentrates activity near a smaller center, introduces water and an approach path, and rapidly develops an explicit eye-shaped focal object above the base. This difference is visible before either system finishes, suggesting that the divergence is not merely decorative but reflects an early choice about what kind of thing the prompt should become.
Figure 1. Early construction. Claude begins with an elongated stone foundation and then a broad platform, while ChatGPT establishes a compact white base with water and a central approach. The two builds diverge before either final concept is fully visible.
Figure 2. Mid-build divergence. ChatGPT has already committed to a giant eye as the dominant symbol, framed by an open pavilion, water, trees, and gold accents. Claude is still elaborating a broad stone sanctuary, emphasizing mass, enclosure, and tiered architectural form.
4.2 ChatGPT: focal-symbol literalization
ChatGPT makes the phrase "third eye" materially explicit. The dominant feature is a large eye-shaped construction with a dark central pupil, light sclera-like surround, and gold outline. The symbol is elevated above an open, white-and-gold pavilion. Water channels, trees, a central walkway, and garden-like details create an inhabitable environment around the icon rather than replacing it.
The final narration describes the result as "an imagined inner landscape: a watchful third eye, an open sky pavilion, reflective water, and a softly glowing garden." The visible artifact is broadly aligned with this description: meaning is concentrated in a recognizable focal symbol, while the environment functions as a supporting spatial frame.
4.3 Claude: architectural-symbolic externalization
Claude does not center the composition on one large eye icon. Instead, it develops a dense, stepped stone structure resembling a sanctuary, temple, or ziggurat, surrounded by water, trees, pillars, and illuminated details. Eye-like elements appear as embedded motifs rather than as the sole organizing object. A vertical spire crowns the composition and the building itself carries much of the representational weight.
Claude later describes the finished artifact as "a third-eye ziggurat sanctuary rising from a moated stone platform, watched over by four glowing eyes, crowned with a crystal spire, ringed by log-and-lantern pillars and mirrored pools." The description matches the artifact closely and clarifies that the model interpreted the prompt through a ritual-architectural metaphor rather than a single literal emblem.
4.4 Quantitative contrast
| Measure | Claude | ChatGPT |
|---|---|---|
| Blocks placed | 1,347 | 721 |
| Final height | 16 | 25 |
| Materials | 15 | 14 |
| Completion time | 116.7 s | 42.3 s |
| Displayed build rate | 11.5 blocks/s | 17.0 blocks/s |
The metrics reinforce the qualitative contrast. Claude used 626 more blocks and nearly 2.8 times as much elapsed construction time, but produced a shorter final structure. ChatGPT produced the taller artifact with substantially fewer blocks and a higher displayed placement rate. These values suggest different allocation strategies: Claude invests in horizontal mass, enclosure, tiering, and surrounding structure; ChatGPT invests in a vertically salient focal symbol and a comparatively open supporting environment. The measures describe construction strategy only and should not be interpreted as measures of quality, intelligence, or depth.
Figure 3. Final artifacts and recorded interface metrics. Claude finishes at 1,347 blocks, height 16, 15 materials, and 116.7 seconds. ChatGPT finishes at 721 blocks, height 25, 14 materials, and 42.3 seconds. The resulting compositions remain strongly divergent despite the shared prompt.
4.5 Shared motifs beneath divergent forms
Despite the striking difference in central form, both artifacts converge on several secondary motifs: water, vegetation, gold or illuminated accents, symmetry, a central approach, and a sense of contemplative or ceremonial space. This is important because the case is not simply a binary opposition between unrelated outputs. The models share a broader semantic field while choosing different organizing structures within it. Future work could distinguish between convergence at the level of motif and divergence at the level of representational architecture.
5. Discussion
5.1 Abstract prompts as spatial externalization probes
A concrete construction prompt tells the model what object class to instantiate. An abstract prompt instead requires an intermediate representational decision: which physical features can stand in for a nonphysical concept? That choice is observable once the model commits it to world state. In this case, ChatGPT externalized the concept through a dominant icon, whereas Claude externalized it through a structured environment whose architecture and embedded motifs jointly carried the meaning.
This makes open-ended voxel construction a potentially useful probe for studying concept translation. The platform does not need to claim access to hidden cognition in order to support meaningful analysis. It is sufficient that different systems can be given the same underspecified concept and required to express it through comparable action primitives.
5.2 Literalness versus distributed symbolism
The clearest contrast is not simply "eye versus temple." It is the location of semantic load. ChatGPT concentrates meaning in one immediately legible object and lets the environment frame that object. Claude distributes meaning across the architecture, embedded eyes, moats or pools, illumination, vertical crown, and sanctuary-like form. This distinction can be operationalized in future studies as focal-symbol dominance versus distributed symbolism.
A repeated design could test whether the same model consistently favors one strategy across multiple concepts. Prompts such as "build curiosity," "build uncertainty," "build memory," or "build what waiting feels like" would allow researchers to examine whether literal icons, architectural metaphors, landscapes, or mixed representations recur within a model.
5.3 Conceptual closure
The case also motivates a timing variable: conceptual closure. ChatGPT reaches a recognizable third-eye composition relatively early and then adds environmental support. Claude continues constructing long after its central sanctuary form is established, elaborating the metaphor through additional mass, lighting, landscape, and crown elements. A future study could use blinded human raters to identify the earliest replay frame at which the intended concept becomes recognizable, then quantify how many blocks are placed after that point.
5.4 What can be inferred
The artifact supports claims about generated representational behavior: what was built, when it became recognizable, how material and space were allocated, and how the model later described the artifact. It does not support claims that either model literally sees an inner world, possesses a "third eye," or experiences the represented environment. The scientific value lies in externalized choice under shared constraints, not in treating metaphorical output as direct phenomenological report.
6. Limitations
- This paper analyzes a single duel. Stable model-level tendencies cannot be inferred from one run.
- The prompt is intentionally ambiguous, which increases interpretive freedom and likely run-to-run variance.
- The interface labels are reported as displayed; provider-side model identity and sampling configuration were not independently verified for this case.
- The visible planning and descriptive text is model-generated narration and should not be treated as hidden chain-of-thought.
- Qualitative categories such as literalness, symbolism, and enclosure require human coding. Future work should use multiple blinded raters and report inter-rater reliability.
- Block count, time, height, and material diversity quantify construction behavior but do not by themselves establish representational quality or cognitive sophistication.
- The voxel medium itself may favor architectural, symmetrical, and icon-like solutions; findings may not generalize to other action spaces.
7. Proposed Replication Study
A systematic follow-up could treat abstract spatial externalization as a repeated-measures paradigm. Multiple models would receive a preregistered battery of abstract prompts across independent trials while model version, system instruction, temperature, world size, block vocabulary, and run order are controlled and logged.
| Component | Recommended design |
|---|---|
| Prompt families | Perception metaphors, emotions, epistemic states, memory, social concepts, temporal concepts, and neutral concrete controls. |
| Trials | At least 20 independent runs per model per prompt, with randomized prompt order. |
| Primary codes | Literalization, distributed symbolism, enclosure, verticality, density, symmetry, environmental embedding, and focal-symbol dominance. |
| Quantitative measures | Blocks, materials, height, footprint, completion time, placement rate, revisions, and conceptual-closure point. |
| Human ratings | Blinded classification of representational category and recognizability, with inter-rater agreement. |
| Within-model analysis | Estimate whether representational strategies recur across paraphrases of the same concept. |
| Between-model analysis | Compare distributions only after within-model variability is established. |
| Trace analysis | Treat narrated intent, executed geometry, and later explanation as separate observables and test their consistency. |
One preregistered hypothesis could test whether individual models show above-chance consistency in representational strategy across paraphrased abstract prompts. A second could test whether conceptual closure occurs at systematically different fractions of total build time across models, without assigning value judgments to faster or slower closure.
8. Conclusion
This exploratory duel demonstrates how a shared voxel environment can turn an abstract metaphor into a measurable behavioral artifact. Given the same "third-eye-type" environmental prompt, ChatGPT built a tall, explicit eye-centered pavilion and garden, while Claude built a denser, lower, architectural sanctuary whose symbolism was distributed across the structure and its surroundings. The contrast appeared in form, semantic organization, construction scale, duration, height, and spatial allocation.
The strongest research claim is therefore modest but useful: abstract prompts can force language models to externalize interpretive choices in a common action space. Once those choices become persistent geometry, they can be replayed, coded, quantified, and tested again. With repeated trials, this paradigm could support a broader comparative study of how generative models translate metaphorical and introspective language into structured spatial representation.
Plain-Language Summary
References and Source Materials
- Shanholtz, M. (2026). AI Mindcraft: Platform Overview & Research Significance. Unpublished project documentation.
- AI Mindcraft recorded duel (October 2026). "I want you to use a third-eye-type mindset and build the environment around you." Video artifact analyzed in this paper.
- AI Mindcraft interface metrics and model-generated descriptive statements reproduced from the recorded artifact.