What makes a teaching prompt produce something usable
Generic prompts get generic worksheets. The fix is to load four specifics: the exact grade or level, the learning objective or standard, the time you actually have (a 45-minute period is not a unit), and the classroom reality (class size, range of readiness, available materials). Add an output shape — a table, a numbered plan, a rubric grid — and the model produces something you can drop into your planner.
Two guardrails matter for accuracy. First, ask the model to mark anything it's unsure of rather than inventing a fact, date, or statistic — it will fabricate plausible-sounding 'facts' for content if you let it, so verify any subject-matter claim before teaching it. Second, keep student data out: describe students by role or need ('a student reading two grades below level'), never by name or record.
These prompts run well on any current model. For high-volume routine work (quiz questions, vocabulary lists, email drafts), an efficiency-tier model like Gemini 3.1 Flash-Lite or gpt-5.4-mini is plenty; for nuanced differentiation and feedback, a frontier model reasons better. Prices as of June 2026 (OpenAI, Gemini).