OverviewWhat it is
Prompt engineering designs the input - instructions, context, examples, output format - to reliably get the output you want. It is the cheapest, fastest lever, and always the first thing to try before RAG or fine-tuning.
At a glancePrompt Engineering
Prompting is the cheapest, fastest lever; always try it before RAG or fine-tuning.
CompareThe prompting-technique landscape, side by side
Every technique worth reaching for, grouped by the kind of move it is — basics, ways to elicit reasoning, structure and control, and tool/retrieval augmentation. Filter or search; each row shows what it looks like in a prompt.
Every row has a page — what it does, what it costs you, and how to tell when it is the thing biting you.
MechanicsHow it works
Techniques run simplest to most powerful: zero-shot (just ask), few-shot (add examples), chain-of-thought (ask for reasoning), and ReAct (reason plus call tools). Structured-output prompting forces JSON or a schema for reliable downstream use.
Ground levelWhat you actually build
A ladder of techniques, each rung more capable and more expensive. It stays the cheapest lever only while the prompt is versioned and tested like anything else you ship.
LandscapeTypes & approaches
Click a highlighted type to open its own page — concept, use case, and diagram.
FeasibilityArchitecture & feasibility
Architecture & feasibility
- Prompts are versioned assets - templating, testing, and a prompt registry belong in the pipeline, not scattered in code.
- Longer prompts (more examples, more context) cost more tokens and latency; the architecture trade-off is quality vs price per call.
- Structured output + validation is what makes an LLM safe to wire into downstream systems.
In practiceWhat it means for building
Huge quality gains for near-zero cost. Prompt iteration is often the fastest path from a shaky demo to something shippable.
Treat prompts as tested, versioned artifacts with eval coverage; a prompt change is a deploy and deserves the same rigor.
GlossaryKey terms
CheckCheck your understanding
How do you improve a bad LLM output?
Iterate the prompt first: clarify the instruction, add few-shot examples, specify the format, try chain-of-thought - before RAG or fine-tuning.
Zero-shot vs few-shot?
Zero-shot just asks; few-shot adds examples for reliability on specific formats, at the cost of a longer, pricier prompt.
How do you make prompts production-safe?
Version them, force structured output, validate it, and cover them with evals so changes are tested.
What changedWhat changed here
- Introducing ChatGPT Images 2.5
ChatGPT Images 2.5 with Sketch lets you doodle directly in ChatGPT to guide image generation, making it easier to iterate on visual ideas without prompt engineering.
Three kinds of claim, strongest first. Signal runs every morning.