Computer Science Achievement and Writing Skills Predict Vibe Coding Proficiency

Synthesis: In a preregistered cross-sectional study (N = 100 tertiary students), both written-communication proficiency (r = .29) and computer-science achievement (r = .39) significantly predicted GUI-oriented "vibe coding" performance, with CS achievement remaining a significant predictor after controlling for domain-general cognitive skills — a finding that links prose skill to Large Language Models (LLMs)-driven development; in a joint model CS achievement contributed roughly twice the unique variance of writing skills, though both added independent predictive value, and prompt quality mediated the link from writing skill to vibe-coding success. Findings speak to tool and curriculum design — when to emphasize prompt-writing versus CS fundamentals.

Relevance to AI in Education: This study reframes writing as a primary programming skill in the era of LLM-native software creation, connecting directly to Prompt Engineering research, the construct of vibe coding, CS Education curriculum redesign, and the predictors of who succeeds when AI handles implementation. It provides controlled, preregistered evidence that complements the practice-oriented vibe-coding papers already in this knowledge base.

Key Findings

• Both writing and CS achievement predict vibe coding. In a preregistered, cross-sectional study (N = 100 tertiary students), written-communication proficiency correlated r = .29 with vibe-coding performance and CS achievement r = .39; both were significant independent predictors in a joint model, where CS achievement contributed roughly twice the unique variance of writing skill.
• CS achievement survives controlling for general cognitive ability. CS achievement remained a significant predictor (partial r = .281, p = .005) of vibe-coding performance after controlling for domain-general cognitive skills (ICAR16), indicating the relationship is not merely an artifact of general reasoning.
• Writing works through prompt quality. Human-graded prompt quality mediated the association between writing skill and vibe-coding performance — response-process evidence that clear, structured prose translates into better natural-language prompts, supporting the construct validity of the measure and the importance of written communication in LLM-native development.
• CS knowledge helps even in "no-code" mode (decomposition and algorithmic thinking are facets of Computational Thinking). Because the environment hid the generated source, CS achievement could only operate indirectly (problem decomposition, mental models of control flow); the authors argue this makes their CS estimate a lower bound for general LLM-augmented programming where code editing is also available.
• The pattern extends beyond GUI surface tasks. In the data-centric, constraint-satisfying meal-planning subtask, both CS achievement (r = .320) and writing (r = .202) remained significant predictors, tentatively suggesting the findings are not limited to surface-level GUI construction.
• Exploratory: prior LLM usage negatively correlated with writing and vibe-coding performance (though not with CS achievement) — the authors speculate LLMs may blunt expression, or weaker writers self-select into heavier LLM use; not causal.

What this means for practice

• Instructors. Grade the prompts students submit, not only the artifact: human-graded prompt quality mediated the link between writing skill and vibe-coding performance, and written-communication proficiency remained an independent predictor (r = .29).
• Curriculum designers. Keep CS fundamentals in the curriculum even for no-code workflows — CS achievement was the stronger predictor (r = .39; partial r = .281 controlling for ICAR16) and operated only indirectly when the generated source was hidden, making its estimated contribution a lower bound.
• Instructors. Use source-hidden, GUI-oriented tasks when you intend to assess expressive specification; in mixed-mode environments where students may read or edit generated code, expect CS knowledge to carry more of the load.
• Curriculum designers. Add written-communication instruction to software courses, because decomposition and algorithmic thinking must pass through prose when students specify behavior to an LLM.
• Researchers. Report prior LLM usage alongside performance: it correlated negatively with writing and vibe-coding performance (though not with CS achievement), a non-causal pattern consistent either with LLM use blunting expression or with weaker writers self-selecting into heavier use.

Limitations

• N = 100 tertiary students recruited from a single laboratory volunteer pool (ETH Zurich and University of Zurich), mean age 25.0 (SD = 3.5), 51 of them engineering or technology majors, each paid 55 CHF for one 1-hour-45-minute session; all had completed an introductory CS course and had prior experience using LLMs for programming tasks, so novices were excluded by design.
• The written-communication instrument was newly designed for this study and had never been deployed before, which the authors list as a construct-validity threat; essay graders needed recalibration and rescoring after an ICC(2,2) of 0.735 fell below the preregistered 0.750 threshold.
• The study assessed "pure" no-code, GUI-oriented vibe coding in a controlled laboratory with the generated source hidden, so results may not transfer to permissive LLM-augmented workflows that expose code, to non-GUI tasks such as data analysis, or to professional developers and citizen programmers.
• Tests and tasks were strictly timed — several participants ran out of time while close to a correct solution — and the cross-sectional design supports prediction, not causal claims about what improves vibe-coding skill.

Citation

Thorgeirsson, S., Weidmann, T. B., & Su, Z. (2026). Computer Science Achievement and Writing Skills Predict Vibe Coding Proficiency. Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems (CHI '26), Barcelona, Spain. 17 pages.