Keep prompts short; reasoning degrades fast

2026 best-practice writeups note LLM reasoning quality starts dropping around ~3,000 tokens, with a practical sweet spot of 150–300 words for most tasks. The field is splitting cleanly into casual prompting vs. production context engineering — treat them as different skills.

---

Sources: Thomas Wiegold — Prompt Engineering Best Practices 2026