The techniques that reliably help
Being specific about format, length, and audience beats a vague request every time — "write a 3-bullet summary for a non-technical manager" produces a more directly usable answer than "summarize this," because the model no longer has to guess what output shape you actually want.
Giving the model a small number of examples of the exact input/output pattern you want ("few-shot prompting") is one of the most reliable levers there is — it works because the model is very good at continuing a pattern it's just been shown, better than it is at inferring an unstated one.
Getting the model to reason before it answers
Asking a model to "think step by step" before giving a final answer (chain-of-thought prompting) measurably improves accuracy on tasks with multiple logical steps — math, multi-step reasoning, anything where jumping straight to a conclusion skips work the model actually needs to do.
For tasks with a genuinely correct answer format, giving the model a way to check its own work — "before finalizing, verify each step against X" — catches a real share of errors the first pass would have missed, the same way a human proofreading their own work catches things a first draft didn't.
What prompt engineering can't fix
No amount of clever phrasing gets a model to reliably know something it was never trained on or has no way to look up — a prompt can shape how the model answers, not manufacture facts it doesn't have access to; that's a job for retrieval (RAG), not prompting.
A prompt that works well on one model version isn't guaranteed to work identically on the next one — models get updated, and prompts tuned against very specific quirks of one version are the most likely to break; the techniques above (be specific, show examples, ask for reasoning) tend to be the ones that survive model upgrades, because they're not exploiting a quirk, they're just clearer instructions.