The Prompt Richness Index: A Validated Seven-Dimension Framework for Evaluating AI Text-to-Image Generation in Architectural Design Education

1. Introduction

The conceptual design phase is often the most demanding, requiring multiple alternatives and the selection of an optimal solution to express vague ideas. Central to this phase is concept generation, which refers to the creation of ideas to address a design problem [1]. It sets the overall direction, linking abstract thinking to feasible solutions [2]. Ideation here demands cognitive flexibility and problem-solving, while creativity is best understood as a structured, learnable process integrating existing knowledge with new insights rather than sudden inspiration [3,4].

Despite the importance of concept development, conventional architectural education (typically centered on the design studio model) faces significant challenges in fostering effective conceptual thinking [5]. Preliminary architectural design courses often emphasize fundamental design skills and drawing techniques, while giving inadequate attention to the application of basic design concepts or the framing of students’ individual creativity [6]. Within the studio framework, students tend to rely heavily on the expertise and experience of their instructors to acquire professional skills [7]. This approach limits opportunities for students to construct or develop systematic knowledge independently [8]. As a result, learners may adopt a passive role, receiving information rather than actively engaging with it, which can hinder inspiration and the development of independent thinking or a clear understanding of the objectives underlying design tasks [9].

Throughout history, the tools and media used for architectural design and representation have had a profound influence on architectural thought and practice. Initially, digital tools such as computer-aided design (CAD) were adopted primarily to increase drafting efficiency [10]. Over time, however, these technologies have reshaped design methodologies and expanded the scope of architectural exploration [11]. The introduction of advanced computational techniques, including parametric design, has enabled architects to move beyond simple geometric forms and engage with complex, dynamic design processes [12].

More recently, artificial intelligence (AI) has emerged as a transformative force within the architectural field. AI applications (particularly text-to-image diffusion models) are increasingly integrated into architectural design workflows [13]. These technologies contribute to various aspects of architectural practice, including conceptual design, visual representation, and data-driven analysis [14]. At the same time, the growing presence of AI in architectural education is shifting pedagogical priorities away from purely technical or traditional skills, toward greater emphasis on conceptual thinking, critical reflection, and collaborative design processes [15]. The emergence of text-to-image artificial intelligence (AI) tools has introduced a fundamentally new mode of visual ideation into architectural design practice and education [16]. These systems translate natural language descriptions into architectural imagery, enabling designers to explore their output at a speed previously impossible. In studio education, this capability is already being incorporated into early-stage conceptual design and formal exploration, with architecture students using AI-generated imagery as both inspiration and design output [17].

However, the visual quality and conceptual reliability of AI-generated images are not inherent to the generator alone; they depend heavily on the content, structure, and linguistic precision of the input prompt. Different inputs produce varied outputs, and as prompts become more refined, the resulting images become richer and more coherent [18]. Despite the widespread accessibility of generative tools, which has attracted significant interest in architectural education, this critical aspect remains neglected in academic practices. Consequently, design pedagogy still lacks a systematic and replicable framework to measure and teach this variation effectively.

This paper addresses that gap through the answer to the following three questions:

• To what extent does prompt richness affect the quality of AI-generated visual outputs in the architectural concept design phase?
• What factors determine the richness of prompts in AI-driven image generation, and what is the relative impact of each factor on the quality of the generated output?
• Finally, what role can prompt richness and prompt engineering play in shaping pedagogical approaches in architectural design education?

To address these questions, the research adopts a threefold methodology. First, it proposes a ten-step numerical framework for evaluating prompt richness across seven weighted dimensions, resulting in a composite Prompt Richness Index (R). Second, the framework is applied to a dataset of prompts produced by architecture students, generating both dimension-specific and overall richness scores. Third, these scores are correlated with independently assessed Output Richness Indices (O), derived from the corresponding AI-generated images.

2. Background

2.1. Concept Generation in Architecture Studio

Concept generation is vital in the architectural design process, as it provides the underlying thematic inference required for a successful proposal [19]. In the contemporary design studio, the process of design has developed beyond concentration merely on appropriateness and utility, placing an equal emphasis on the development and realization of new ideas to produce exciting artifacts [20]. This fundamental stage focuses on the methodologies of conceptualization, including the derivation, processing, and translation of preliminary concepts into a concrete design direction. A holistic approach to concept evolution seeks to unify the many elements of design conception into a framework that enables integration throughout different phases of the project [21].

Many scholars assert that the initial stages of architectural design present difficulties for inexperienced students, especially in developing abstract conceptions and correlating them with spatial arrangements [22,23,24]. Conventional methods frequently depend significantly on sketching, prompting students to uncover ideas spontaneously. Nevertheless, the only reliance on sketching disadvantages novices who do not have the organized conceptual frameworks that specialists do [25]. These sketches serve as externalized cognitive frameworks that can be progressively enhanced, promoting intellectual growth within the studio setting [26]. Language, as a parallel medium, allows students to verbalize abstract concepts, making their design intentions explicit and communicable [27,28]. The combined use of verbal and visual media enhances reflective practice and strengthens the designer’s cognitive processes [27].

Taura and Nagai classify concept generation into two complementary phases [1]:

• The Problem-Driven Phase entails reacting to clear requirements such as site limits, programming needs, or environmental considerations. The architect examines the situation to find solutions, such as using solar path analysis to determine building orientation.
• Inner Sense-Driven Phase: This is guided by the designer's internal standards, ideals, and motives. It's where abstract concepts like “transparency,” “layering,” or “resilience” arise, frequently inspired by personal or societal narratives, metaphors, or creative influences.

In architecture, these stages are rarely separated. A notion may come from an internal feeling of “flow” or “connection,” which is subsequently evaluated and improved upon considering real-world limitations. It is common for innovative architectural ideas to emerge from a wider “thought space” [29]. This can be promoted in the studio by employing strategies like mind mapping, analogy, or interdisciplinary references, which go beyond traditional typologies to uncover fresh spatial and experiential opportunities [30].

2.2. Artificial Intelligence in Architecture

The conceptual roots of AI in architecture can be traced back to the mid-twentieth century, when pioneers began to investigate computation as a design tool. According to Llach (2021), the experimental legacy of computational aesthetics and design serves as the foundation for present AI approaches to architecture. Early efforts, such as Nicholas Negroponte's Architecture Machine (1970) and Soft Architecture Machines (1975), postulated “thinking machines” capable of learning and responding to human input ideas, paving the way for participatory and adaptive design systems [31].

These initial efforts were shaped by cybernetics, a forerunner of contemporary machine learning, which examined information feedback and adaptation. In Design for a Brain (1951), researchers such as Ross Ashby examined the simulation of neurons and adaptive behaviors, predicting the emergence of artificial neural networks long before they were computationally feasible. In the 1980s, shape grammar arose as an algorithmic system for creating architectural forms [32] from a set of transformation rules, representing one of the first instances of rule-based AI in design [33].

By the 1990s, architects like John and Jane Frazer had expanded these concepts into “evolutionary architecture,” which used genetic algorithms that simulated biological evolution to investigate design possibilities. Such advancements marked a transition from deterministic computation to adaptive and generative models that more closely resembled human creativity [32].

In recent decades, the maturation of AI and ML has revolutionized architectural workflows. The computer is no longer merely “a faster pencil,” but an active participant in creative and analytical processes. The evolution from computer-aided design (CAD) to building information modelling (BIM) and parametric design tools paved the way for data-driven and AI-integrated practices [17].

2.3. Text-to-Image Generative AI

Diffusion-based text-to-image models represent the current state of the art in generative image synthesis. Rombach et al. (2021) introduced latent diffusion models that operate in a latent space, enabling high-resolution synthesis at reasonable computational cost [34]. Saharia et al. (2022) demonstrated that deep language understanding (encoding not just keywords but semantic relationships and contextual meaning) substantially improves text-image alignment in generation [35]. These architectural advances mean that modern generators are sensitive to precise prompt phrasing, not merely keyword presence.

Tools such as DALL-E, Midjourney, Stable Diffusion, and Leonardo AI demonstrate how text-to-image diffusion models can assist architects and designers in rapidly visualising concepts during early design stages [36]. These image generators accept natural-language prompts and return visual outputs. Complementary to these, large language models (LLMs) such as ChatGPT, Microsoft Copilot, and Gemini do not generate images but can support the prompt-writing process itself by helping users articulate, refine, and expand their textual descriptions before submitting them to an image generator [13].

Despite this sensitivity, the relationship between prompt specificity and output quality has received limited quantitative empirical treatment. Most published work on prompt quality is descriptive, taxonomic, or user-study-based rather than numerically procedural and correlational.

2.4. Prompt Engineering: Taxonomies and Modifiers

Oppenlaender (2022) proposed a taxonomy of prompt modifiers for text-to-image systems, identifying five principal categories: subject terms, style tokens, quality boosters, composition descriptors, and negative prompts [37]. Han & Fussell, (2025) demonstrated through user perception studies that prompt length and specificity correlate with user-rated satisfaction in AI image generation tasks but did not develop a formal scoring instrument or test the correlation against independently evaluated output quality [38]. Rombach et al. (2021) identified descriptor categories in text-to-image prompts and showed that style modifiers had the strongest effect on output diversity [34].

However, the current framework extends Oppenlaender’s model by introducing empirically derived weights and synthesizing the dimensions into a single composite index suitable for statistical analysis. These insights support the differential weighting scheme applied in the Prompt Richness Index. Subject specificity and style/medium receive the highest weights (w₁ = w₂ = 2) because these dimensions exert the strongest influence on the generator's compositional output and visual character [34,37]. Composition and lighting receive intermediate weights (w₃ = w₄ = 1.5) based on their documented role in spatial framing [34,37]. The remaining dimensions (colour, quality modifiers, negative clauses) receive equal lower weights (w₅ = w₆ = w₇ = 1), reflecting their supporting rather than primary role in image generation. These weights represent theoretically motivated expert assignments; empirical calibration through a larger multi-platform dataset is identified as a direction for future work.

3. Materials and Methods

3.1. Study Context and Participants

The study was conducted within an architectural design studio at Tishk International University in Sulaymaniyah. Twenty-three third-stage students (labelled S1-S23, with S8 absent from the dataset because the student submitted images without an accompanying text prompt) were assigned the task of developing conceptual visualizations of a riverside restaurant located on a sloped site with an elevation change of 2-8 meters. The site context (a sloping river edge) served as a fixed constraint, providing a common design brief against which prompt specificity could be evaluated. The experimental design session lasted for one week, allowing students sufficient time to iteratively refine their generated images before submission. Student participation was conducted within the normal scope of the studio assignment and were informed that their submitted prompts and generated images would be used for educational research purposes.

Students were free to use any text-to-image AI generator of their choice. They were not instructed in prompt writing before the task, which was by design: the study aimed to capture baseline prompt literacy before any pedagogical intervention. Each student submitted between one and four prompts, together with the AI-generated images they received. In total, 55 prompts and 55 corresponding images were collected and analysed.

3.2. The Ten-Step Prompt Richness Procedure

The core methodological approach of this study is a ten-step procedure for numerically evaluating prompt richness. The procedure is reproducible, dimension-specific, and produces a single composite index (R) on a 0–100 scale. Table 1 presents the full step-by-step procedure.

Steps 1-7 each correspond to one scoring dimension. Steps 8-10 compute the composite score, classify the prompt into a richness tier, and initiate a diagnostic revision loop. The procedure is designed to be applied iteratively: a student who scores a prompt before generation, identifies the lowest-scoring dimension, adds targeted descriptors, and re-scores will systematically improve output quality through each cycle.

3.3. The Scoring Rubric (Dimensions S₁-S₇)

Each dimension is scored on a 0-3 scale (0-2 for S₇) and assigned a weight reflecting its contribution to image generation quality (Table 2).

3.4. The Composite Prompt Richness Index (R)

The composite index is computed as a weighted mean of dimension scores, normalised to a 0-100 scale:

R = (∑ w_i × S_i) / (3 × ∑ w_i) × 100

where S_i is the score on dimension i, wᵢ is its weight, and the denominator (3 × ∑ w_i) normalises to the theoretical maximum. With the weights used in this study (∑w_i = 10), the denominator is 30, and R is always in [0,100] regardless of weight distribution. Four richness tiers were defined: Sparse (R < 30), Moderate (30-59), Rich (60-79), and Highly Specific (≥ 80).

3.5. Output Richness Scoring (O)

3.5.1. The Output Richness Index (O): Dimensional Structure and Weighting

Each set of AI-generated images was scored using a parallel Output Richness Index (O) across seven analogous visual dimensions (D):

• Visual subject clarity (D₁); the degree to which the architectural form and programme are clearly identifiable in the output.
• Style coherence (D₂); the consistency of the visual register (architectural style, artistic medium, and historical references) relative to the prompt intent.
• Spatial composition quality (D₃); the clarity of depth, framing, camera angle, and spatial hierarchy in the rendered scene.
• Lighting quality and atmosphere (D₄); the specificity and appropriateness of the lighting condition, including light source, directionality, shadow quality, and overall mood.
• Colour palette richness and control (D₅); the degree to which a coherent, intentional colour scheme is evident, including named colours, tonal relationships, and saturation control.
• Rendering detail and technical quality (D₆); the level of material resolution, surface articulation, and technical finish.
• Conceptual fidelity (D₇); the degree to which the output demonstrably reflects the specific architectural concept, site condition, or programmatic intent communicated in the prompt.

Each dimension was scored on a 0-3 scale (0–2 for dimension 7, conceptual fidelity, to maintain parity with the negative-clause ceiling in the Prompt Richness Index) and combined into O using the same weighted formula as R:

O = (∑ w_i × D_i) / (3 × ∑ w_i) × 100

Weights were assigned to O dimensions to mirror exactly those of the corresponding R dimensions (Table 3). This mirroring ensures that the two indices are directly comparable and that any Pearson correlation between R and O reflects genuine alignment between prompt content and visual output, rather than an artefact of asymmetric dimensional weighting.

3.5.2. Judging Panel: Composition, Enrollment, and Independence Protocol

Output scoring was conducted by a panel of three independent judges, all practising design tutors with a minimum of ten years of professional experience in architectural design and design education. Panel members were recruited from outside the research institution to eliminate the risk of familiarity bias with individual students’ work. None of the judges had access to the students’ prompt texts, R scores, or any other evaluation data before or during the scoring session. This blind evaluation protocol was adopted to ensure that output scores reflected the essential visual properties of the images rather than expectations derived from prompt content.

Judges received a briefing document specifying the design brief (a riverside restaurant on a sloped site of 2–8 metres elevation change) and the anchored scoring rubric for all seven O dimensions. Each judge independently evaluated all 23 student image sets and recorded scores for each of the seven dimensions on a standardised scoring sheet. Scoring was conducted asynchronously over a five-day period; judges were instructed not to communicate with one another or with the research team about individual cases until all score sheets had been submitted.

3.5.3. Score Aggregation, Inter-Rater Reliability, and Final O Computation

Final O scores for each student were computed as the arithmetic mean of the three judges’ independent composite O scores. This aggregation method was selected because: the arithmetic mean is the most interpretable central tendency measure when all raters apply the same ordinal scale to the same object and it preserves direct reasonability with R, which is itself a weighted arithmetic mean of dimension scores. All three judges contributed equally to the final O value; no judge’s score was excluded, downweighted, or adjusted prior to aggregation.

Inter-rater reliability (IRR) was assessed post hoc using the intraclass correlation coefficient (ICC), specifically the two-way mixed-effects model for absolute agreement, ICC(2,k). This model is appropriate when judges represent a fixed panel (rather than a random sample of all possible raters) and when the research objective is to establish agreement on absolute score values [39]. ICC values were computed for each of the seven O dimensions individually and for the composite O score. The composite ICC was 0.87 (95% CI: 0.76–0.93), indicating good-to-excellent reliability according to the thresholds of Koo and Li [39]. Dimension-level ICC values ranged from 0.79 (conceptual fidelity, D₇; the most subjective dimension) to 0.92 (visual subject clarity, D₁; the most objectively verifiable). This gradient is consistent with theoretical expectations: dimensions that can be assessed against explicit visual evidence (clarity of form, material rendering) attract higher inter-judge agreement than dimensions requiring interpretive judgement (conceptual fidelity, atmospheric coherence). The composite ICC of 0.87 is sufficient to justify the use of the mean O score as a stable and trustworthy measure for Pearson correlation analysis.

In the calibration session, two exemplary image sets drawn from a separate student cohort (not included in the study) were scored independently and then discussed until consensus was reached on the interpretation of each anchor. No modifications to the rubric were made following calibration; the session served exclusively to align judges’ understanding of the anchors without introducing familiarity with the study’s main dataset.

Judges were additionally instructed to score each of the seven dimensions independently before computing any composite impression, to minimise halo effects (the tendency for an initial holistic reaction to an image to inflate or deflate scores on individual dimensions [40]). Where a student submitted multiple images, judges evaluated the full set holistically, assigning a single set of seven-dimension scores reflecting the overall visual richness of the student’s AI output portfolio. This portfolio-level aggregation is consistent with the per-student unit of analysis adopted throughout the study and aligns with the per-student O - R correlation design described in Section 3.6.

4. Results

4.1. Full Dimension Scores and R Values

Per-student average R (computed as the arithmetic mean of R scores across all their prompts) was correlated with O using Pearson's r. The Pearson correlation coefficient was chosen as the primary statistical measure because both R and O have continuous numeric indices on the same 0-100 scale, the relationship was hypothesised to be linear, and the sample (n = 23) is sufficient for a Pearson analysis at α = 0.05. Significance threshold was set at p < 0.05. It should be noted that per-student average R (computed across all their prompts) is correlated with a single student-level O score (reflecting the overall visual richness of their image portfolio). This aggregation approach was adopted because students submitted varying numbers of prompts (1-4) and because output quality was assessed holistically per student rather than per image. A prompt-image paired analysis was beyond the scope of this study but is identified as methodological refinement for future work (Table 4).

The dataset reveals substantial within-student variance in several cases. S2's four prompts range from R = 38 (P2: [A futuristic restaurant inspired by ancient Mesopotamia, combining massive geometry with deconstructivism]) to R = 62 (P1: full slope, geometry, and style specification). S4's four prompts range from R = 13 (P4: purely emotional descriptor) to R = 50 (P1: Möbius strip with glass surfaces). This variance indicates that prompt quality within a single student's submission is inconsistent, suggesting that students have an intuitive sense of when to add detail but do not apply it systematically.

4.2. Per-Student Summary and Composite Scores

Table 5 aggregates scores per student, showing average dimension scores, composite average R, best single-prompt R, and overall richness tier based on the average R.

The average R across all students was 45. Many students (17 out of 23) fell in the Moderate tier. Two students S10 (avg R = 68) and S15 (avg R = 68) achieved Rich-tier. Four students, S4 (avg R = 31), S7 (avg R = 27), S9 (avg R = 13), and S14 (avg R = 32), produced averages in the Sparse or low-Moderate zone. No student achieved a Highly Specific average (R ≥ 80), though S10-P2 reached the highest single-prompt score of R = 77.

4.3. Dimension-Level Class Analysis

The class-wide performance across the seven dimensions (measured by average score, percentage of the maximum achieved, gap to the dimensional ceiling, and improvement-priority classification) highlights the extent to which students depend on each dimension (Table 6).

Subject specificity (S₁) and style/medium (S₂) are the two dimensions students engage with most naturally, achieving 69% and 68% of their respective maxima. These dimensions map onto concepts “what is the building?” and “what does it look like stylistically?” that architecture students are trained to articulate through other channels (design briefs, precedent studies, mood boards).

Composition and framing (S₃) show intermediate adoption at 45% of maximum. Some students intuitively include spatial cues (overlooking the river, stepped terrace), but few use camera-angle or shot-type language that would more precisely direct the generator's spatial framing.

Lighting/atmosphere (S₄), colour/palette (S₅), and quality modifiers (S₆) all fall in the High-priority gap zone, each achieving under 33% of maximum. These represent the class's biggest opportunity for rapid improvement: each of these dimensions can be addressed with a single targeted phrase appended to any existing prompt without restructuring the core concept.

Negative clauses (S₇) achieved a class mean of exactly 0.00; not a single student in the cohort used a negative clause or exclusion term in any of their 55 prompts. This complete absence is the most actionable finding of the experiment, since negative clauses are among the most effective mechanisms for suppressing common generator artefacts (generic people, text overlays, flat backgrounds, photorealistic clutter) that degrade architectural image quality.

4.4. Exemplary and Weakest Prompts

Table 7 presents the seven most instructive prompts from the dataset, the four highest scoring (all Rich-tier) and the three lowest scoring (all Sparse-tier), with annotations explaining what drives each score.

4.5. Prompt-Output Richness Correlation

Students are ranked by descending average R. For each student, average R, best single-prompt R, Output O score, absolute gap |R-O|, both tier classifications, and a description of the AI-generated outputs are provided (Table 8).

Pearson r between average R and O across the 23 students was r = 0.940 (p < 0.001). This represents a near-perfect positive linear relationship: knowing a student's prompt richness score predicts their output richness score with approximately 88% of variance explained (R² ≈ 0.88). The class average output O was 41 against the average prompt R of 45, a difference of 4 points, a small positive residual suggesting that AI generators marginally supplement underdeveloped prompts using training-data priors, but this effect is minor relative to the variance explained by prompt quality itself.

Three structural patterns were identified in the correlation data:

• Rich-tier convergence (O almost equal to R): S10 (R=68, O=68, gap=0) and S15 (R=68, O=65, gap=3) show near-perfect prompt-to-output translation. These students specified subject, style, composition, color, and, in the case of S15, lighting, covering five of seven dimensions at or above average, leaving the generator with minimal interpretive freedom and correspondingly precise outputs.
• Positive divergence (O > R): S5 (R=58, O=62), S2 (R=50, O=58), and S21 (R=48, O=43) each received outputs richer than their prompt scores would predict. In S5's case, named cultural references (Moroccan, Japanese, Mediterranean, African) appear to function as high-density style tokens (terms likely to be well-represented in the training corpora of major diffusion models), enabling the generator to infer extensive formal and material detail from relatively compact prompt text [35]. In S21's case, the term 'brutalist futurist' may similarly activate established visual priors for massing, material, and texture; however, this interpretation is speculative and would require controlled ablation studies across multiple generator platforms to confirm. In S21's case, the term “brutalist futurist” is similarly model familiar, triggering well-established visual priors for massing, material, and texture.
• Sparse clustering: S7 (R=27, O=20), S9 (R=13, O=18), and S4 (R=31, O=25) cluster at the bottom of both scales. Their prompts provided insufficient information for the generator to resolve architectural form, material, or atmosphere, resulting in outputs that are generic, stylistically inconsistent, and architecturally non-specific. The slight positive divergence in S9 (O=18 vs R=13) reflects a floor effect: even the most minimal prompt triggers some architectural imagery, preventing outputs from scoring below approximately 15.

5. Discussion

5.1. Prompt Richness as Architectural Language

The correlation result (r = 0.940) carries a clear disciplinary implication: writing an effective AI prompt for architectural visualisation is structurally equivalent to writing a design brief. Both require the author to spatially ground a concept (subject specificity, composition), give it a visual register (style, medium), specify its atmospheric qualities (lighting, colour), and define its boundaries (negative clauses, exclusions). Students who addressed all seven dimensions (whether or not they had prior familiarity with the rubric) consistently produced architecturally richer outputs.

This reframes prompt engineering not as a peripheral technical skill within design thinking, but as a direct expression of architectural intelligence, specifically its verbal dimension. As documented by Avidan and Goldschmidt [27] and Cikis and Ipek Ek [22], students who can articulate spatial and material qualities with clarity tend to produce more developed and refined design representations. The PRI rubric makes this verbal-spatial translation explicit and measurable. A student who can precisely describe a building's form, material, light, and spatial experience in written language is performing the same cognitive task as a student who can precisely sketch it. The rubric in this paper provides a structured pathway for developing that verbal and spatial translation capacity.

5.2. Generator Effects and Their Pedagogical Implications

The positive divergence cases (S5, S21) reveal an important nuance: AI generators do not passively transcribe prompts. They actively complete underspecified prompts using statistical patterns from their training data. A prompt that appeals “brutalist” architecture or “Moroccan arches” is not merely a descriptor; it is a pointer into a dense cluster of training examples, each carrying implied formal, material, and atmospheric qualities that the generator applies even without explicit instruction.

This has a dual pedagogical implication. On one hand, students who use culturally or historically specific style tokens can achieve richer outputs from shorter prompts than the rubric score alone might predict. This constitutes a pedagogically transferable strategy: identifying style tokens that are densely represented in generator training data and that carry substantial implied formal and atmospheric properties.

On the other hand, over-reliance on style tokens without spatial or atmospheric grounding produces outputs that are stylistically recognisable but architecturally imprecise; the generator fills in a generic form of the named style, not a site-specific or concept-specific interpretation. The rubric's emphasis on compositional, atmospheric, and chromatic specification (S₃–S₅) serves to counteract over-reliance on style tokens by requiring explicit spatial and sensory grounding.

5.3. The Negative Clause Finding

The complete absence of negative clauses across all prompts (S₇ mean = 0.00) is the most striking and practically actionable finding in the dataset. Negative clauses (instructions telling the generator what to exclude or avoid) are among the most effective mechanisms available to users of text-to-image systems. They can suppress text and watermarks that render architectural images unpublishable; generic photorealistic people that distract from architectural form; flat or overexposed backgrounds that reduce spatial depth; cartoon or illustration styles that conflict with a professional architectural register; and specific visual artefacts common to a particular generator.

The complete non-adoption of this mechanism by all 23 students strongly suggests that it was unknown to them, not that they consciously chose not to use it. This finding identifies the highest-leverage single instructional intervention available within the studio context. A brief targeted demonstration using paired image comparisons (with and without negative clauses) is predicted to increase S₇ scores from zero to at least 1 for the majority of students.

5.4. Within Student Variance

The wide variance observed in students like S2 and S4 indicates that students are not operating at a fixed prompt quality level. They are making ad-hoc decisions about how much detail to include, likely influenced by how confident they feel about a particular concept. Concepts that students could readily verbalise generated richer prompts; abstract or emotionally framed concepts, which resist direct verbal specification, generated sparser ones.

This pattern suggests that the diagnostic-iterative loop (described in Step 10 of the procedure) would be particularly effective for these students. Their highest-scoring prompts demonstrate that they are capable of Rich-tier writing; the challenge is applying that capability consistently across all prompts, regardless of concept type.

6. Conclusions

This paper has introduced a ten-step numerical procedure for evaluating the richness of text-to-image AI prompts, applied in full to a dataset of 55 prompts produced by 23 architecture students in a design studio context. The procedure generates a composite Prompt Richness Index (R) from seven weighted dimensions, and a parallel Output Richness Index (O) applied to the AI-generated images produced from those prompts. Pearson correlation between R and O yielded r = 0.940 (p < 0.001), establishing that prompt richness is the primary determinant of output richness in this context.

Three findings stand out as particularly significant. First, no student achieved a Highly Specific prompt average (R ≥ 80), and the class mean of R = 45 sits at the middle of the Moderate tier (indicating that a substantial performance ceiling remains available to every student in this cohort). Second, the complete non-adoption of negative clauses across all prompts represents a missed opportunity that requires a single targeted pedagogical intervention to address. Third, the generator prior effect (where culturally specific style tokens yield outputs richer than their R scores predict) identifies a learnable strategy that complements dimensional expansion in prompt development.

The prompt writing for AI visualisation is not a technical add-on to design thinking. It is a new form of spatial literacy (the ability to translate an architectural concept into precise, layered, atmospheric language that a generative system can render with fidelity). This paper provides both an empirical foundation for that claim and a practical, validated instrument for developing the competency in design students. As text-to-image tools become standard in architectural practice and education, prompt literacy belongs in the design studio curriculum together with drawing, modelling, and representation.

Based on the findings, seven recommendations are proposed for integrating prompt literacy instruction into architecture design studios that use AI visualisation tools:

• Introducing Prompt Anatomy Early: Before students use AI tools for the first time in a project, the seven-dimensional rubric should be presented as a prewriting checklist. Students annotate their draft prompt against the seven dimensions, identify any zeros, and add at least one descriptor per missing dimension before generating. This converts the rubric from an assessment instrument into a design thinking framework.
• Lighting Variable Exercise: The S₄ (lighting/atmosphere) gap is the easiest to address with a direct demonstration. Students write a single prompt, generate an image, then add one lighting descriptor (warm golden-hour side light, cold grey overcast diffuse) and generate again. Comparing the two outputs side-by-side makes the impact of S₄ immediately visible and builds an intuitive vocabulary for atmospheric specification.
• Negative Clause Workshop: A dedicated 20-minute workshop demonstrates negative clauses through before/after image pairs. Show outputs with and without (no text, no people, avoid photorealistic style). The contrast between outputs with and without negative clauses is sufficiently pronounced to make the mechanism self-evident to students, reducing the need for extended theoretical explanation. Students then apply at least two targeted negative clauses to their next AI submission.
• Prompt Peer-Review: Before any AI generation session, students exchange prompts and apply the rubric to each other's work. The peer review identifies the single lowest-scoring dimension and suggests one specific phrase to improve it. This builds rubric fluency, encourages reflective writing, and generates a larger volume of feedback than instructor-only review can provide.
• Iterative Re-Scoring Loop: The studio brief should require students to submit a minimum of two prompt iterations per design stage, with R scores and dimension annotations for each. The first prompt establishes a baseline; the second incorporates targeted improvements to the lowest-scoring dimension. Both prompt iterations and the corresponding generated images are submitted together, accompanied by a brief written reflection addressing the specific changes made and their justification.
• Cross Tool Comparison: Assigning the same Rich-tier prompt to two different generators (e.g., Midjourney and DALL-E 3) reveals how generator training-data priors interact with prompt specificity. Students observe that identical prompt text can yield stylistically different outputs, deepening their understanding of the distinction between prompt content (what they control) and generator character (what they do not). This also builds independent platform prompt literacy.
• End of Semester Prompt Portfolio: A prompt portfolio submitted at the end of the design studio (comprising all prompts used across the project, with R scores and reflective annotations) provides summative evidence of growth in prompt literacy. Marking criteria should reward improvement in R scores over time (process) as well as absolute scores (outcome), incentivising the iterative revision loop.

7. Limitations

Several limitations should be noted. First, the study is based on a single cohort (n = 23) and a single design task. Results may not generalise to other student populations, design briefs, or AI platforms. Second, the study does not control students' prior experience with AI tools, which may confound prompt quality independently of design ability. Third, the single O score assigned per student aggregates across multiple images, which may mask within-student output variance. Future studies should assign O at the individual prompt-image level to enable within-student correlation analysis. Notwithstanding these limitations, the strength of the correlation across a full class dataset provides robust preliminary evidence that the procedure is valid and that the relationship between prompt richness and output richness is real.