Prompt engineering

Zero-shot prompting states the task without examples. It relies entirely on what the model already knows how to do. It is the default form of every ordinary request.

Few-shot prompting supplies example outputs inside the input to show what is wanted. The examples clarify the format more than the content. It is the cheapest way to fix an output shape.

Chain-of-thought prompting asks for explicit step-by-step reasoning before the answer. It improves performance on tasks with several steps. It also makes the reasoning inspectable.

Context engineering manages everything in the context that is not the prompt: system instructions, memory, retrieved documents and tools. It is the wider discipline the prompt sits inside. Most failures attributed to prompts are failures of context.

Adversarial prompting tests safety by deliberately trying to provoke unwanted behaviour. It is evaluation rather than attack. It is how limits are discovered before somebody else discovers them.

Tokenization splits input into the sub-word units a model consumes. It determines cost and effective length. It also explains why some phrasings behave unexpectedly.

Top-k sampling restricts the next token to the k most probable candidates. It limits how far the model can stray. It is one of two standard truncation methods.

System instructions define the operational boundaries and character of a model's responses. They sit above the user's message. They are also the target of most attacks.

Prompt injection is input crafted to override the original instructions. It exploits the fact that instructions and data arrive in the same channel. It is a structural vulnerability rather than a bug.

Prompt leaking is tricking a model into revealing its confidential system instructions. It exposes what an operator intended to keep private. It is a recognised and common failure.

Jailbreaking is bypassing safety filters through crafted input. It is the adversarial counterpart to alignment. The arms race between the two is continuous.

Tree of thoughts explores several branching reasoning paths and selects among them. It generalises chain-of-thought from a line to a tree. It costs more and solves harder problems.

Retrieval-augmented generation adds retrieved domain knowledge to the prompt. It is how a model answers about material it never saw in training. It is the main architectural pattern in production.

In-context learning is a model adapting to a task purely from what is in the prompt, without any weight change. It is temporary and immediate. It is why examples work at all.

Model parameter tuning adjusts inference settings such as temperature to control output variation. It shapes behaviour without touching the model. It is often more effective than rewriting a prompt.

An automatic prompt engineer uses a model to generate and evaluate candidate prompts. It replaces human iteration with search. It routinely finds prompts a person would not write.

It covers readable text: instructions, examples, reasoning steps. It leaves out the parameter-level techniques that steer models without words, the sampling methods that decide what comes out, the indirect attack surface, and the measurement of when a model refuses too much or too little.

Prefix tuning trains continuous vectors prepended to the input rather than composing text. It steers a model without any readable prompt. Discussions focus on manual text composition.

Self-consistency decoding samples several reasoning paths and takes the majority answer. It converts a single guess into a vote. Explanations centre on single-path output.

Indirect prompt injection places the attack in content the model will read later, such as a web page or a document. The user never types it. Security analyses focus on direct input only, which misses the more dangerous case.

Graph of thoughts allows reasoning steps to connect in arbitrary ways rather than in a line or a tree. It permits merging and revisiting. Strategies stop at tree structures.

Directional stimulus prompting uses a small auxiliary model to generate hints that steer the main one. It adds guidance without adding context. Techniques emphasise direct insertion instead.

Skeleton-of-thought generates an outline first and expands its parts in parallel. It reduces latency on long outputs. Workflows assume sequential generation.

Soft prompting uses learned continuous vectors in place of discrete tokens. It is prompt engineering below the level of language. Guides restrict definitions to text.

Self-refine has a model critique and improve its own output across iterations. It moves effort from the prompt to the loop. Design focuses on the initial construction.

Nucleus sampling selects from the smallest set of tokens whose probabilities exceed a threshold. It adapts the candidate set to the situation, where top-k fixes it. Parameter discussions omit it.

Models use material at the start and end of a long prompt more reliably than the middle. Tutorials assume uniform attention. Where you place an instruction changes whether it is followed.

Sycophancy is a model agreeing with a premise the user embedded in the question. Evaluations assume impartial reasoning. A leading question reliably produces a leading answer.

Prompt distillation compresses a long working prompt into a shorter one with the same effect. It reduces cost per call. Tutorials focus on expanding prompts.

Red teaming is systematic adversarial testing rather than ad-hoc attempts. It is a procedure with coverage goals. Security testing is usually described as improvisation.

Refusal rate measures how often a model declines to answer. Safety studies focus on identifying malicious triggers. The rate itself is the operational number.

Over-refusal is declining harmless requests. Filtering analyses assume a binary boundary. A system that refuses too much fails its users while looking safe.

Prompt inversion reconstructs a hidden system prompt from observed outputs. It is the analytic counterpart to leaking. Security models focus on output safety instead.

Differential privacy in prompting applies formal guarantees to limit what can be inferred about data included in a prompt. Privacy is usually treated qualitatively. The formal version is available and unused.

Your page is read as part of somebody's prompt, alongside instructions you never see and other documents competing for the same space. Content that states its point early, stands alone, and cannot be misread survives that assembly. Everything else is edited out before the reader arrives.