#prompting

What immediately caught my eye is the premise itself: that there’s a “load-bearing vocabulary” in Claude. That sounds trivial on first read, but I think it points at something practical that people building with these models run into all the time: some prompts work not because they’re conceptually special, but because the model has learned to treat certain words as signals of seriousness, structure, or intent. The annoying part is that this can be true even when nobody involved can explain exact

Meta-prompting is the practice of using one prompt to design, improve, or control another prompt, rather than asking the model to do the task directly. It helps when a plain prompt is too vague, too brittle, or too hard to maintain. Instead of hand-tuning a long instruction by trial and error, you can ask an LLM to generate, refine, critique, or transform prompts for you. In practice, teams reach for meta-prompting when they want: better prompt quality with less manual iteration, reusable prompt

Tree-of-thought prompting, often abbreviated ToT, is a prompting method where an LLM explores multiple reasoning branches instead of committing to a single straight-line answer. Tree-of-thought prompting is useful when a problem benefits from deliberate search: puzzle solving, planning, multi-step reasoning, or tasks where an early mistake can derail the whole answer. In practice, it gives the model a way to "think in branches," then compare candidates before choosing one. You'd reach for To

Self-consistency is a way to get a language model to answer by generating several reasoning paths and then choosing the answer that appears most often, instead of trusting a single chain of thought. A single reasoning trace from an LLM can be brittle: one unlucky step can derail the final answer. Self-consistency helps when the task needs multi-step reasoning, such as math word problems, logic puzzles, or multi-hop question answering. In practice, you reach for it when you want a cheap reliabili

Using XML tags or other delimiters in prompts means wrapping parts of your prompt in clear markers like `<context>...</context>` so the model can separate instructions, data, and examples more reliably. This pattern helps when a prompt contains multiple kinds of content at once: instructions, source text, user input, examples, or fields you want the model to treat differently. Delimiters reduce ambiguity and make prompts easier to read, debug, and reuse. In practice, teams reach for XML-style ta

A prompt template is a reusable prompt with placeholders for variable details, so you can generate consistent LLM inputs without rewriting the whole prompt each time. Prompt templates solve a simple but common problem: most prompts are partly fixed and partly changing. In practice, they help you: keep instructions consistent across many requests swap in user-specific data, documents, or task details reduce copy-paste errors when building apps, workflows, or evals make prompts easier to version,

Structured output, often called JSON mode, is a way to ask an LLM to return data in a fixed machine-readable format instead of free-form text. When you want an answer that another program can safely parse, plain chat text is fragile. A model might add extra words, change field names, or wrap the result in an explanation. Structured output helps when you need: predictable API responses reliable extraction into downstream systems cleaner automation for forms, workflows, or tools outputs that match

Prompt engineering is the practice of designing and refining the text you give to a language model so it produces more useful, reliable, and specific outputs. Large language models are sensitive to how you ask. A vague request can produce a vague answer; a well-structured prompt can improve format, scope, tone, and task performance. You reach for prompt engineering when you want to: get a model to follow instructions more consistently, steer it toward a role, style, or output format, reduce back

A system prompt is the instruction text that sets the assistant’s role, behavior, and boundaries before the user’s message is processed. System prompts help you control how an LLM behaves without changing the model itself. They’re useful when you want consistent tone, safety rules, formatting, or task-specific behavior across many requests. In practice, teams use system prompts to: keep outputs on-brand, enforce policies or guardrails, define what the assistant should and should not do, reduce a

Few-shot prompting, also called in-context learning, is a way of asking a language model to do a task by giving it a few example inputs and outputs in the prompt. Few-shot prompting helps when you want the model to follow a pattern, format, or classification rule without training or fine-tuning a new model. You’d reach for it when: the task is simple but the output needs to be consistent, you can show the model 1–5 good examples, you want to prototype quickly before investing in fine-tuning

Zero-shot prompting is asking an AI model to do a task without giving it any worked examples first. It is the simplest way to use a large language model: you describe the task in plain language and let the model infer the pattern from its pretraining. That makes it useful when you want: a fast baseline before writing better prompts simple classification, extraction, or drafting tasks lower prompt length and less example curation a test of whether the model already “gets” the task well enough In

Chain-of-thought prompting, often abbreviated CoT, is a way of asking a language model to produce intermediate reasoning steps before giving the final answer. It is used to improve performance on tasks that benefit from multi-step reasoning, such as arithmetic, logic, and some question-answering problems. In practice, teams reach for CoT when a direct one-line answer is more likely to be wrong than a stepwise solution. It can also make model behavior easier to inspect: if the model shows its

Context engineering is the practice of deliberately assembling the right information, instructions, tools, and examples into an AI model’s context so it can do a task reliably. Large language models do not “know” what you want unless you provide it in the prompt, surrounding messages, retrieved documents, tool outputs, or examples. Context engineering helps you get better accuracy, fewer hallucinations, and more consistent behavior by shaping what the model sees before it answers. In practice, i

A context window is the amount of text a language model can read and keep in mind at once while it generates a response. The context window sets the practical limit for what you can ask a model to consider in a single turn: your prompt, system instructions, chat history, retrieved documents, and the model’s own recent output all compete for that space. In practice, it matters when you want the model to: follow long instructions, answer questions about a large document, maintain conversation stat

Prompt caching is a way to reuse the cost and latency of processing the same prompt prefix instead of sending or recomputing it from scratch every time. If your app sends a long, repeated system prompt, policy block, tool schema, or document context on many requests, you can end up paying to re-process the same text over and over. Prompt caching helps reduce that waste. In practice, teams reach for it when: a large part of the prompt stays the same across requests, they want lower latency for re

Prompt injection is when untrusted text causes an AI model or agent to follow attacker-controlled instructions instead of the user’s intended task. Prompt injection matters because modern LLM apps often mix trusted instructions with untrusted content: web pages, emails, documents, tickets, chat messages, or tool outputs. If the model treats that content as instructions, it can be tricked into leaking data, ignoring policies, or taking unintended actions. You usually care about prompt injection w

Jailbreaking is the act of tricking a model or device into bypassing its normal safety, policy, or security restrictions. For AI systems, jailbreaking is usually about getting a model to ignore refusal rules, reveal restricted information, or produce disallowed content. For devices and operating systems, “jailbreaking” means removing vendor-imposed limits so you can install unapproved software or gain deeper system access. In practice, the AI meaning is the one most developers mean today. It mat

steering は、LLMの内部表現（activations）を直接いじって、出力の傾向を変える手法 たとえば「簡潔に答えて」といった性格を、プロンプトではなくモデル内部から強められる ただし、軽い調整は prompting で代替できることが多く、万能ではない 一方で、拒否を弱める・検閲を外す など、prompting では難しい用途がある Open weights の強いモデルが増えたことで、ローカル環境で steering を試す価値が急に上がった ただし著者は、実用面ではまだ懐疑的。結局は fine-tune や prompt のほうが効率的では、と見ている この記事のテーマは、ざっくり言うと 「LLMの出力を、プロンプトではなく内部から直接いじる技術」 です。 これが steering。日本語にするなら「誘導」や「操舵」に近い感じでしょうか。 著者 Sean Goedecke は、昔の「Golden Gate Claude」にかなり魅了されたと言います。これは Claude の出力がなぜかゴールデンゲー