Secondary analysis links prompt specificity to musical role in 24 users

Text-to-music systems can turn a short description into a complete musical clip, lowering barriers to composition while changing who makes the consequential creative decisions. A secondary analysis of 304 prompts submitted by 24 users in the PAGURI study found that people with different self-described musical roles encoded markedly different amounts of technical-musical information in their prompts, but did not differ in how consistently they refined one prompt into the next. The exploratory result, published in Frontiers in Psychology, separates the apparent ability to articulate a musical intention from the interaction process through which that intention may be shaped.

The researchers analysed the publicly available ordered prompt corpus from PAGURI, a one-hour user-experience study built around AudioLDM 2 and a DreamBooth-inspired personalisation method. The sample included listeners, dance teachers, performers, conservatory and ensemble members, solo artists, DJs, producers and technical practitioners. Participants had a mean age of 26.9 years, most were Italian, musical engagement was generally high and prior experience with text-to-music systems was limited. The analysis used the participant, rather than each of the 304 prompts, as the primary statistical unit, avoiding the false independence that would result from treating repeated prompts from the same person as separate cases.

To measure how explicitly users described music, the team counted instrument and voice terms alongside technical parameters such as tempo, key, chords, reverb, delay, distortion, compression, attack and release. A second measure, specificity density, expressed the number of such markers per 10 words so that differences were not simply a consequence of writing longer prompts. The researchers then created three exploratory groups from self-described roles: six producer or technical participants, 10 performers, and eight people in a non-performer or other category. The scheme was explicitly treated as a rough contextual grouping rather than a validated scale of musical expertise.

Producer and technical participants averaged 1.14 musical-specification markers per prompt, compared with 0.84 among performers and 0.19 among non-performer or other participants. The between-group difference was substantial: a Kruskal–Wallis test gave H(2) = 10.25, p = 0.006, with an epsilon-squared effect size of 0.39. The pattern persisted when prompt length was taken into account. Specificity density averaged 2.13 markers per 10 words for producer and technical participants, 1.42 for performers and 0.33 for the other group; the corresponding test yielded H(2) = 11.52, p = 0.003, with epsilon-squared = 0.45. Mean prompt length itself did not differ significantly between the groups, with H(2) = 1.08 and p = 0.583, so the group differences were not explained simply by one group writing longer prompts.

That difference in what people said did not translate into a measurable difference in how they steered the system. The study calculated Jaccard similarity between the content-word sets of consecutive prompts, treating common function words as noise, and checked the pattern with TF-IDF cosine similarity. At a descriptive Jaccard cut-off of 0.20, 131 of 280 consecutive pairs, or 46.8%, were classified as refinements and 149, or 53.2%, as restarts. Participant-average refinement rate was 0.463, with a standard deviation of 0.211 and a bootstrap 95% confidence interval of 0.383 to 0.548. Rates were nearly identical across role groups: 0.472 for non-performer or other participants, 0.475 for performers and 0.432 for producer and technical participants. The group comparison was not significant, H(2) = 0.24, p = 0.888, while prompt specificity was not associated with refinement rate, with Spearman’s rho = 0.050 and p = 0.818.

Continuous similarity measures told the same story, avoiding reliance on a single arbitrary dividing point. Mean Jaccard similarity across all 280 pairs was 0.261, with a standard deviation of 0.309 and median of 0.162, while participant-level TF-IDF continuity did not vary across role groups, with H(2) = 0.04 and p = 0.980. The two similarity measures were almost perfectly rank-correlated, with Spearman’s rho of 0.986 and p less than 0.001. Their agreement shows that the broad result was not an artefact of one calculation, although it does not make either metric a direct measure of conceptual refinement: users may pursue the same musical goal with substantially different wording, or retain similar wording without meaningfully refining the goal.

The balance between refinements and restarts also depended on where the threshold was placed. Using cut-offs from 0.10 to 0.30, the proportion classified as refinements fell from 56.8% to 50.4%, 46.8%, 42.5% and 36.1%, respectively. Restarts formed a majority only at thresholds of 0.20 and above, so the researchers treated continuous similarity as the primary continuity measure and the 0.20 cut-off only as an interpretive summary. The data therefore support the restrained conclusion that lexical continuity was modest overall and broadly similar across the role groups, rather than the stronger claim that users generally preferred to restart. A sensitivity analysis that reclassified the ambiguous role description “Play music non professionally” from non-performer/other to performer also preserved the central pattern: the specificity differences weakened but remained statistically significant, and refinement rates still did not differ across groups.

Repeated interaction also did not produce a common upward trend in technical specificity. Within-person correlations between iteration number and specificity could be estimated for 21 of the 24 participants; three had constant specificity throughout their sessions. The mean correlation was −0.06, with a standard deviation of 0.44, eight positive slopes and 13 negative slopes, and the distribution did not differ significantly from zero, with Wilcoxon W = 89.0 and p = 0.374. A comparison of the first three and final three prompts produced a small, non-significant decline from 0.82 to 0.69 markers per prompt, with W = 66.5 and p = 0.407. This does not show that users learned nothing: the index captured explicit technical vocabulary, not semantic, evaluative or workflow learning, and one session may be too short to reveal durable change.

From these behavioural patterns, Zhang and Pan propose a tentative framework for perceived creative control, defined as the subjective sense of directing consequential choices in a creative process and its outcome. The prompt corpus did not directly measure that construct, so the framework was not empirically tested by the analysis. It separates three design conditions: task structure, or how many meaningful decisions the task asks users to make; system autonomy, including whether an interface offers local steering or simply produces a complete result; and model coverage, meaning whether the underlying system can represent the intended musical tradition or grammar. Prior musical experience is treated as a potential moderator. The authors argue that technical vocabulary alone is not a measure of creative control: a participant may know how to name instruments, tempo or effects yet still lack a way to enforce those intentions through successive generations.

The resulting propositions are hypotheses rather than demonstrated educational benefits. The authors predict that decision-rich tasks can increase perceived control when system autonomy is held constant, while highly autonomous systems may reduce psychological ownership if they remove consequential human choices. Interfaces offering local revision, parameter locking or persistent steering are predicted to support more control than repeated whole-output generation. Perceived control is proposed as a possible mediator between engagement conditions and creative self-efficacy, psychological ownership and sustained engagement, while weak model coverage is expected to limit the benefits of better steering. These claims require direct testing with experimental manipulations, validated psychological scales and substantially larger samples; the prompt corpus contained no item-level measures of perceived control, self-efficacy or ownership.

The findings carry important limits for music education. The source study was not a classroom intervention: it included no teachers, instructional tasks or learning outcomes, and nearly 95% of participants were Italian, with 62.5% master’s-degree students, mainly in music and acoustic engineering. The role labels were self-described, the specificity dictionaries were not validated against expert coding, and protocol-imposed personalisation pseudowords may have mechanically increased similarity between prompts. All behaviour came from one model and interface configuration, limiting generalisation to newer systems. Nevertheless, the dissociation offers a concrete design question for future controlled testing: whether generative music tools should optimise for polished output alone, or preserve consequential choices through section-level regeneration, editable structure, visible parameter controls and clear documentation of what the model can and cannot realise. The answer requires experimental, psychometric, cross-cultural and longitudinal studies before it can guide practice with confidence.

Subject of Research: Secondary analysis of prompting behaviour in AI-assisted music creation

Article Title: Prompt specificity and iterative behaviour in text-to-music co-creation: a secondary analysis of interaction logs and a framework for perceived creative control

Article References: Zhang, X., & Pan, L. (2026). Prompt specificity and iterative behaviour in text-to-music co-creation: a secondary analysis of interaction logs and a framework for perceived creative control. Frontiers in Psychology, 17, Article 1965465. https://doi.org/10.3389/fpsyg.2026.1965465

Image Credits: AI Generated

DOI: 10.3389/fpsyg.2026.1965465

Keywords: text-to-music, generative AI, prompt specificity, iterative continuity, creative control, human–AI co-creation, music education, interaction logs, psychological ownership, PAGURI study, Prompt, specificity