Prompt Engineering
The craft of writing instructions that get the best, most reliable output from an AI model.
Reviewed July 10, 2026Stable
When not to use it
- As a substitute for evaluation. Prompt tweaks feel productive and prove nothing without a test set. "It looks better" is how teams ship regressions.
- When the real problem is data or model choice. No prompt rescues a model that has never seen your domain, or a task that needs a database lookup.
- At scale, as a permanent fix. Prompts that carry heavy instructions on every call cost money on every call. At volume, fine-tuning or a smaller model is cheaper.
Reach for something else instead
- Few-shot examples usually beat elaborate instructions. Show, don't explain.
- Structured output constraints (schemas, grammars) are more reliable than asking politely for JSON.
- Fine-tuning once a prompt has grown to hundreds of tokens of rules you repeat every request.
Read more on the blog
- Prompt engineering in 2026: what still works and what changedIf your prompts still open with "act as an expert" and "let's think step by step," you're using 2023 advice on 2026 models, and some of it now makes your output worse. What actually moves results today, what quietly stopped working, and why.
- Why AI answers the question you asked, not the one you meantAsk a model whether it can do something and it may tell you rather than doing it. The gap between literal meaning and intended meaning has a name, a fifty-year literature, and a pattern in where models fail.
- Why AI aces the test and fails the variationTransformers scored 96 to 99% on a semantic parsing benchmark and 16 to 35% on the same task with the pieces recombined. The gap has a name, a 35-year argument behind it, and a fix that suggests the capability was there all along.
- RAG vs fine-tuning: the decision, honestlyThe most common question in applied AI, answered without the hedging: when do you use retrieval, when do you fine-tune, when do you need both, and what almost every team gets wrong about the choice.
Further reading
- Brown et al. (2020), Language Models are Few-Shot Learners — where few-shot in-context learning was demonstrated at scale and named. Its own paper credits GPT-2 (Radford et al., 2019) with showing zero-shot task transfer first.
- Wei et al. (2022), Chain-of-Thought Prompting Elicits Reasoning in Large Language Models.
- Zhao et al. (2021), Calibrate Before Use — few-shot results swing wildly on example order, which is why prompt tweaks need measurement, not vibes.
Primary sources, listed so you can check the claims on this page rather than take them on trust.
Where people go wrong
- Cargo-culting phrases like "you are an expert" or "think step by step" without measuring whether they help for your task on your model.
- Over-instructing. Long prompts full of edge cases often perform worse than short ones with good examples.
- Assuming a prompt transfers between models. It frequently doesn't, and the failure is quiet.
At a glance
FieldLanguage & LLMs
Core ideainstructions shape output
Key techniquesexamples, step-by-step, explicit format
Try beforefine-tuning
DifficultyBeginner
Often compared with
Prompt Engineering vs. Fine-tuning — steering with instructions vs. changing the model itself.
Where this sits
24 concepts come first. Understanding it opens up 16 more.
11Levelsteps in
24Needs firstconcepts
16Opens up5% of 310
4Areascrosses fields
Learn these firstLarge Language Model
Computed from the prerequisite graph, not assigned. How this works