A framework rooted in cognitive psychology and neuroaesthetics can tell an AI agent exactly what separates scroll-stopping content from the forgettable majority.
Cognitive psychology · neuroaesthetics · agent workflows. A working rubric for the machines now making the content.
Scroll through any brand's Instagram feed and you will notice the same gravitational pull toward the obvious: the clean flat-lay, the motivational quote in a sans-serif font, the before-and-after that announces its thesis before you finish blinking. The content is competent. It is also, in the precise technical sense, unoriginal.
That pull is not laziness. It is arithmetic. Research on divergent thinking shows that the first idea a person generates in response to any prompt carries an average occurrence rate of 47 percent across a sample population, meaning roughly half of all people facing the same problem arrive at the same answer first S10. Guilford's Alternative Uses Test, designed in 1967 to measure creative divergence, found that participants need to generate approximately nine responses before reaching ideas that fewer than ten percent of others would produce S10. Originality lives past the comfort zone of the first draft.
The implications compound for AI systems. A language or image model trained on the aggregate of human production is, by construction, a weighted average of what people have already made. Ask it for a post about wellness and it reaches for the statistical mode: soft light, neutral palette, affirmation copy. The model does exactly what its training optimizes it to do, which is to predict the most probable output. Divergent thinking research measures this same tendency in humans and calls it low fluency combined with low originality S9. Guilford identified four components of creative divergence: fluency, originality, flexibility, and elaboration S9 S10. An AI agent running a single-pass generation optimizes for fluency, producing one response quickly, while originality, flexibility, and elaboration go unmeasured and therefore unprompted.
One critical caveat: divergent thinking scores predict creative potential, not creative performance S11. A high score on a controlled test does not guarantee a scroll-stopping image, and the lab conditions of the Alternative Uses Test differ substantially from the real friction of platform algorithms and audience attention. The research establishes a directional truth, not a formula.
That directional truth is still useful. If first ideas cluster at 47 percent commonality, and novelty requires pushing past that initial cognitive basin S10, then any workflow that accepts the first generated output is structurally biased toward the forgettable. The question is what framework forces the push.
Seven theories of creativity, developed independently across six decades, converge on a single uncomfortable claim: creativity is not a gift distributed unevenly among the fortunate few. It is a structure with identifiable components, and those components can be specified, measured, and trained.
Guilford saw it first. His Structure of Intellect model broke creative performance into divergent production operations and transformation products, with fluency, flexibility, originality, and elaboration as the operative dimensions S16. The insight was architectural: creativity depended not on a single faculty but on a multiplication of independent factors working together S16. Boden extended this structural logic from individuals to conceptual spaces, arguing that the most powerful creative acts are deliberate restructurings of the rules that define what a domain considers intelligible S14. Transformational creativity, in her framing, modifies the axioms of a space and cascades change through every dependent rule S14. Guilford mapped the mind; Boden mapped the idea-space the mind moves through.
Amabile pulled the frame outward to include social context, establishing that intrinsic motivation and domain expertise interact with creative thinking skills to produce observable output. Martindale supplied a neurological mechanism, showing that creative cognition correlates with broad, defocused attention, a flat arousal state that allows distant associations to surface before convergent judgment prunes them. These two perspectives bracket the same phenomenon: the creative act requires both the right internal conditions and the right cognitive gear.
Berlyne approached the problem from the audience's side. His empirical aesthetics identified arousal potential as the engine of aesthetic preference; stimuli that are novel, complex, or uncertain generate interest precisely because they require resolution S15. Too little arousal potential produces boredom. Too much produces avoidance. The sweet spot, where attention engages and lingers, is measurable in principle and targetable in practice. Ramachandran and Hirstein then specified the neural mechanisms underneath that curve, showing that artists who apply peak shift, perceptual grouping, symmetry, and visual metaphor are not following intuition but exploiting circuits that evolution built for survival and that reward systems reinforce S12. Exaggerated stimuli activate visual brain areas more strongly than naturalistic ones S12; this is wiring, not preference.
The Geneplore model, developed by Finke, Ward, and Smith, closes the loop operationally. It describes creativity as a cycle: generate a preinventive structure, a rough, ambiguous form; explore it for interpretable properties; then generate again with those findings loaded. The output of one pass becomes the input of the next. Creativity, in this model, is inherently iterative. Accepting the first generated structure is not efficiency; it is premature termination.
Lay these seven frameworks against each other and the agreements become precise. Guilford and Geneplore both insist on iteration over initial output. Boden and Amabile both locate creativity in the modification of constraints rather than their abandonment. Berlyne and Ramachandran both ground aesthetic response in measurable neural and perceptual mechanisms. Martindale and Guilford both require the breadth of associative reach before the depth of evaluative focus. The convergence is not coincidental. These researchers worked from different disciplines and methods but kept arriving at the same structural core: creativity requires novelty that is nonetheless interpretable, surprise that lands within a perceptual system equipped to receive it, and transformation that respects enough of the prior space to remain meaningful S14 S12 S19.
That shared structure underwrites the six principles that follow. Originality draws on Guilford's infrequency criterion and Boden's transformational logic. Expressiveness reflects Martindale's arousal theory and Amabile's motivational conditions. Aesthetic Appeal operationalizes Ramachandran and Hirstein's perceptual laws directly S12. Technical Execution connects to the hierarchical visual processing that Berlyne's model assumes the viewer brings to any encounter with an image S15. Unexpected Associations trace to Martindale's defocused attention and the peak shift mechanism S12. Interpretability and Depth answer the Geneplore requirement that preinventive structures resolve into something a viewer can actually read. Each principle is a lever on a mechanism these seven bodies of work collectively describe.
The six principles operate simultaneously as a scoring rubric, a set of generative instructions, and a diagnostic vocabulary. Each maps to at least two of the seven theories and translates that theoretical grounding into a specific demand on the image. Scoring runs from 1 to 5. A composite threshold of 3.5 is a working heuristic drawn from a single study context, not a validated benchmark; treat it as a starting point and calibrate it against your own engagement data.
The six principles carry equal theoretical weight, but equal weight is the wrong instruction for an agent building content for a specific platform. A score that treats Instagram and LinkedIn as interchangeable environments will produce work that is technically adequate everywhere and resonant nowhere. The weightings proposed below are expert judgement, not empirically calibrated coefficients. Practitioners must validate them against their own engagement data before treating the composite score as a reliable instrument.
Across all four platforms, the principles remain constant; their relative pull changes with the audience's attention posture and the platform's structural mechanics. An agent running the same brief across all four should produce four distinct creative solutions, each weighted toward the principles that the platform's own dynamics make most decisive.
The template's value collapses immediately if the agent generating content never receives precise enough instructions to apply it. Prompt vagueness is the first failure mode, and it is structural. As one practitioner account of AI-assisted design put it, the people who win in these workflows are not those who master the tools but those who "can articulate precisely enough how you want something" S2. Precision here is the mechanism by which rubric criteria enter the generation process at all.
Stage one is the generative brief. Before the agent produces anything, the brief must encode three things: the platform's dominant principle weighting from the previous section, a concrete creative constraint for each of the six principles, and a target register that specifies tone rather than just visual style. The brief should name what the image must accomplish on each dimension: the Originality threshold it needs to clear, the emotional texture Expressiveness should carry, the compositional logic that will satisfy Aesthetic Appeal. Prompt wording for aesthetic qualities requires iteration because different phrasings activate different regions of a model's learned associations, and the gap between "striking" and "visually tense" as descriptors can produce substantially different outputs S5.
Stage two is multi-variant generation. The agent should produce several distinct visual candidates from the same brief. This is where treating visuals as preinventive structures becomes operational rather than theoretical S1S6. Each candidate is a structure that may contain latent directions worth developing, associations worth extending, formal choices that could be pushed further. The Geneplore model treats generation and exploration as a deliberate loop, not a sequence with a defined endpoint. An agent running this stage should therefore vary its generative assumptions across candidates: different compositional logic, different approaches to the Unexpected Associations criterion, different handling of visual complexity. Variants that feel wrong are often more diagnostically useful than variants that feel adequate.
Stage three is scored evaluation and refinement. Each candidate receives a score on all six principles, and any candidate falling below threshold on a given dimension receives a specific diagnostic instruction rather than a generic rejection. This is where the shift in creative skill matters most. When execution is cheap and fast, the scarce capability becomes the ability to look at an output and say, precisely, what is not good enough and what edit would correct it S2. A diagnostic note for a sub-threshold score on Interpretability and Depth should not read "this image lacks meaning." It should read: "the visual metaphor does not carry sufficient ambiguity; add a secondary element that the viewer must actively resolve." That instruction re-enters the generation loop as a refined brief, and the cycle runs again.
Agent design for this workflow requires treating the agent itself as a primary caller of the rubric, not a passive executor of human judgment S5. The scoring logic, the diagnostic language templates, and the platform-specific weighting should all be embedded in the agent's skill structure, not held externally by the operator. This reduces context loss between evaluation and regeneration S6 and allows the pipeline to run at speed without degrading creative judgment at each pass. The agent handles variant generation and scoring; the human practitioner enters at the diagnostic and strategic layer, where the judgment calls about which direction is worth pursuing cannot yet be fully automated S22.
The three-stage pipeline described above is only as good as the content strategy surrounding it. Wire it into a calendar dominated by performance anxiety and it will systematically produce safe work.
The research here requires caveats about sample quality and self-reporting bias, but the directional finding is consistent enough to take seriously. Nearly half of creators report that algorithm changes increase their stress S17, and approximately 43% say that comparing their results with others damages their sense of achievement S17. The behavioral consequence is predictable: creators lean toward formats that have worked before S17. Creative strategies rarely collapse in a single decision; they erode through small accumulating choices, each individually defensible, until teams are generating predictable ideas by default S17. A scoring rubric applied to that psychology will optimize for the middle of the distribution.
The structural fix is a calendar policy. Some teams already dedicate a recurring slot to experimental content, decoupled from normal performance expectations S17. The Creativity Template makes this practice more rigorous. Experimental slots should be scored primarily on Originality and Unexpected Associations, the two dimensions most likely to be discounted when engagement benchmarks govern every post. Aesthetic Appeal and Technical Execution still apply as floors; a piece should not be rough simply because it is experimental. Engagement figures for these posts should be reviewed in aggregate across several iterations, because daily fluctuations will otherwise generate false signals that kill promising directions before they compound S17. Setting success criteria before publication, rather than reading them off the results, further insulates the experimental cadre from the distortion that real-time metrics introduce S17.
The feedback loop runs in both directions. Template scores predict creative quality; empirical engagement data validates or challenges that prediction over time S18S20. Where high-scoring posts underperform, the rubric weights need re-examination. Where low-scoring posts overperform, the platform weighting logic may be miscalibrated for a specific audience. This calibration loop transforms the template from a general framework into an instrument tuned to actual conditions.
Which brings the proviso that governs the whole system. The composite threshold of 3.5 out of 5.0 is a provisional starting heuristic drawn from the reasoning structure of the framework, not a figure validated against measured engagement outcomes. Treat it as a working hypothesis. Run it, track the correlation between scores and the engagement metrics that matter to your specific objectives, and adjust the threshold accordingly. A rubric that never gets revised against evidence is just an opinion that learned to use numbers.
Twenty-two references across transcripts and the research literature. Citation markers in the text link here.