You should be rewriting your prompts (maxleiter.com)

AI practitioners should routinely rewrite and retune prompts when switching models, not just reuse the same text and expect better results. The author—who’s tested many LLMs quickly after release—argues prompts “overfit” to particular models the way models overfit data: differences in preferred formats, positional weighting, and intrinsic biases mean identical prompts can yield wildly different outcomes. Ignoring this leads to apples‑to‑oranges comparisons, higher costs, and degraded accuracy (the gpt‑5/Cursor rollout and the later fixes are cited as an example). Three concrete technical drivers make prompt rewrites necessary: 1) prompt format — some models favor markdown while others (e.g., Claude 3.5) perform better with XML because of their training data; 2) position bias — models weight prompt sections differently (e.g., Qwen may prefer relevant context at the end while Llama prefers it at the start), and 3) model biases from training, RLHF, and post‑training tweaks that change defaults (censorship, verbosity, hallucination patterns). Practical implications: test and eval prompts for each model, monitor behavior, and adapt prompts or post‑processing to work with a model’s defaults rather than fighting them to reduce cost, improve reliability, and avoid prompt overfitting.