Chat-driven vibe-coding builders converge on a single generate-and-check script whose dominant time cost is dead waiting for the tool¶

Claim · All claims
Evidence · 1 study (2 entries) · associational r1–r2 · q2
Currency · newest study 2026 · not yet checked against its sources · not reviewed by a person

Subclaims¶

q2 i? All seven chat-driven builders' sessions match the canonical generate-and-check loop more often than a shuffled within-person null (conformance 0.21–0.80; z = +2.8 to +6.2), and the strongest recurring motif appears about 1.89 times more often than a first-order Markov null predicts. → Léonard Boussioux* 2026 q2 i? Generation intervals cover 119 minutes, a quarter of recorded time, and 61% of generation time is dead waiting. → Léonard Boussioux* 2026 (2)

Evidence¶

Léonard Boussioux* 2026¶

Léonard Boussioux, Ziyi Zhao, and Kanghyun Cho*. (2026). One Tool, One Taste? How Vibe Coding Trades Collective Diversity for Individual Creativity. arXiv. https://arxiv.org/abs/2609.38183 doi:10.48550/arxiv.2609.38183

q2 · i? · associational · r2

Process-layer analysis of nine screen-recorded think-aloud build sessions coded at one-second resolution (28,818 coded seconds; 54 prompt cycles). Loop conformance was tested against a within-person null of 1,000 shuffled sequences; motif mining used a per-person first-order Markov null. Between-session pointwise agreement is modest, so the shared object is the loop, not a schedule.

"For all seven chat-driven builders, the actual session matches the canonical loop far more often than the shuffled version does (conformance 0.21–0.80;𝑧=+ 2.8to+6.2across all seven)."

Léonard Boussioux* 2026 (2)¶

Léonard Boussioux, Ziyi Zhao, and Kanghyun Cho*. (2026). One Tool, One Taste? How Vibe Coding Trades Collective Diversity for Individual Creativity. arXiv. https://arxiv.org/abs/2609.38183

q2 · i? · associational · r1

Waiting-economics decomposition of generation intervals across the recorded sessions. The article reports 61% dead waiting, 20% peeking, 13% productive fill, and 6% idle, and notes productive fill rises from 9% in the first half of a person's intervals to 17% in the second half.

"Within those intervals, pooled across all generation seconds, 61% is dead waiting (73 minutes total, watching the spinner), 20% is peeking at the preview, 13% is productive fill, and 6% is idle (Figure 11)."

Discussion¶

Related Claims¶

• Originality is almost never invoked in builders' spoken acceptance verdicts, and defects as well as defaults exit the loop by acceptance without verification — related
• Builders' perceived control, satisfaction, and perceived quality show no detectable association with their measured visual atypicality, so convergence goes unnoticed from the inside — related
• Vibe-coded websites built for 73 distinct businesses converge into roughly a dozen recurring visual designs despite an explicit originality incentive — related
• Low comment rewriting does not imply a smooth process: students describe the main effort as verifying generated code rather than revising prompts — related
• A substantial share of builders' voiced design intentions never reaches the typed prompt, leaving design choices to the production stack's defaults — related
• Verbal acceptance with residual doubt from the intuitive-error agent signalled convergence: it occurred in seven of eight sessions, and all seven of those converged — related
• Students' help requests to the AI tutor were predominantly instrumental rather than executive, with procedural content and low verification — related