core-concepts

Asking Claude to comb through a batch of data and flag whatever clearly deviates from the normal range, attaching a possible direction for the cause — this is the act of finding what looks off, and nothing more. It doesn't include judging whether that oddity is actually an error; that judgment gets handed to someone actually capable of making it.

Ask Claude to write out the intermediate reasoning steps within a single reply, instead of jumping straight to the conclusion — you don't just see the answer, you see how it was derived, which means you can also see exactly which step went wrong.

Piling more and more individually reasonable constraints into a single prompt — professional but not stiff, concise but not missing detail — where each one holds true on its own, but past a certain point the combined effect stops adding up to the sum of the parts and starts working against itself instead.

Context Handoff refers to the deliberate technique of transferring work context between different Claude conversations. Because Claude has no cross-conversation memory by default, Context Handoff lets you quickly 'wake up' the needed background when starting a new conversation, avoiding the need to re-explain everything from scratch.

The total amount of information Claude can 'see' in a single conversation — including your messages, uploaded files, and all prior conversation turns. Once exceeded, the earliest content gets cut off.

Custom Instructions are fixed background descriptions you configure in <a href="/en/glossary/claude-tools/claude-projects/" target="_blank">Claude Projects</a>, so Claude automatically knows your role, work context, preferences, and constraints at the start of every conversation in that Project — without you needing to re-explain each time.

The order in which examples appear in a few-shot prompt shapes what Claude ends up imitating most — the example placed last tends to get treated as the one that matters most, not a cosmetic detail but an instruction in its own right.

Lets Claude work through a problem in an internal thinking space before giving its formal answer, especially suited to tasks needing multi-step reasoning or prone to mistakes — this thinking isn't necessarily shown to you in full by default; what matters is whether the thinking happened, not whether you got to watch it happen.

Attach two to five concrete input-output example pairs to the instruction, letting Claude pick up the format, tone, and judgment logic by imitation instead of you having to spell every detail out in words — the examples themselves do the explaining.

<a href="/en/glossary/core-concepts/follow-up-question-design/">Follow-up Question Design</a> refers to Claude proactively asking a clarifying question when an instruction isn't specific enough, aiming to confirm direction before actually doing the work, rather than guessing and producing a result that might not match what's needed. Asking isn't a sign of insufficient capability — it's a reasonable response to an ambiguous instruction.

Claude outputs incorrect, fabricated, or unverifiable information with high apparent confidence — like a very confident colleague giving you a wrong answer that feels completely reliable. In workplace contexts, the most dangerous aspect isn't Claude making obvious errors, but making errors confidently.

Iterative Prompting means providing feedback or adjustment requests within the same conversation based on Claude's previous output, progressively moving the output toward what you truly need. Rather than asking once and finishing, it transforms 'Q&A' into 'conversational work' — through several back-and-forth rounds making Claude's output progressively more precise.

"Context" is what Claude can see within a single conversation — it disappears when that conversation ends. "Memory" is information that persists across conversations and is still there next time you open a new chat. Confusing the two leads to broken workflow assumptions.

Model selection trade-off means weighing speed, cost, and capability against each other to pick the Claude model tier that fits the task at hand, rather than always defaulting to the most powerful model or always defaulting to the cheapest.

<a href="/en/glossary/core-concepts/negative-prompting/">Negative Prompting</a> means explicitly telling Claude what not to do, rather than only describing what to do. For example, beyond saying 'write me an apology email,' adding 'don't use stiff corporate language, keep it under 150 words' — those last two instructions are negative prompting.

Output Format refers to explicitly telling Claude in your prompt what the output you want should look like — length, structure, tone, format (bullet points/paragraphs/tables), etc. Explicitly specifying output format is one of the single most effective changes to take Claude's output from 'roughly meets requirements' to 'ready to use directly.'

Break one large task into several separate prompts, where each call's output feeds into the next call's input, chaining together to complete the whole workflow — not one instruction covering everything at once, but staged, with each stage checked before moving to the next.

<a href="https://claude-me.com/en/glossary/prompt-techniques/prompt-injection/" target="_blank">Prompt Injection</a> is when an external document, email, or webpage contains text disguised as an instruction, attempting to trick Claude into executing it as a command from you rather than treating it as plain data to be processed. This matters especially in workplace settings, since you often ask Claude to read documents or emails received from outside sources.

Prompt length calibration means finding the length where a prompt is 'just enough' — too short leaves Claude without enough context to guess what you actually want, too long buries the instructions that actually matter under excess detail. Both ends degrade output quality.

A pre-structured prompt framework with placeholder variables, enabling consistent high-quality outputs by simply swapping in new content each time.

<a href="/en/glossary/core-concepts/response-regeneration/">Response Regeneration</a> means, when you're not satisfied with one of Claude's answers, asking for another answer to the same question without changing the prompt. Since each generation has some randomness, regenerating often produces a version with a different style or angle — but if the underlying issue is that the prompt itself wasn't clear enough, regeneration usually won't fix the root problem.

Role persona depth refers to how much detail is given when setting up a role in <a href="/en/glossary/prompt-techniques/role-prompting/" target="_blank">Role Prompting</a> — giving just a title ('you are a marketing consultant') is a shallow persona, while adding background experience, judgment style, and characteristic tone builds a deep one. Deeper isn't always better; depth needs to match the task's complexity to genuinely improve output quality.

<a href="/en/glossary/prompt-techniques/role-prompting/" target="_blank">Role Prompting</a> is the technique of specifying a particular role or identity for Claude at the start of a prompt — such as 'you are a marketing consultant with 10 years of experience' or 'you are a strict legal editor.' This lets Claude respond from a specific professional perspective, with output that better matches that domain's tone, terminology, and thinking framework.

<a href="/en/glossary/core-concepts/session-continuity/">Session Continuity</a> refers to how different actions within a conversation affect whether Claude still remembers prior context — regenerating a different version of the same message keeps all the conversation before it intact, but opening a brand-new conversation means none of what was previously discussed carries over automatically, unless manually copied over or saved into a Claude Project.

Pulling a small portion — usually ten to twenty percent — of a batch of Claude's output and carefully checking it against the source data, using that small sample's accuracy to estimate roughly how good the whole batch is. The goal isn't catching every single error; it's judging whether the batch as a whole can be trusted.

A <a href="/en/glossary/prompt-techniques/system-prompt/" target="_blank">System Prompt</a> is an instruction pre-configured at the system level before formal conversation begins, defining Claude's role, behavioral guidelines, and response style. What workplace users configure through <a href="/en/glossary/claude-tools/claude-projects/" target="_blank">Claude Projects</a>' Custom Instructions is essentially a form of <a href="https://claudeskill-me.com/en/glossary/prompt-engineering/system-prompt/" target="_blank" rel="noopener">System Prompt</a>.

A <a href="/en/glossary/core-concepts/task-handoff-template/">Task Handoff Template</a> is a fixed-format document that organizes a task's current progress, key decisions made, and unresolved questions, so the next conversation, another person, or your future self can pick up right where it left off without having to review the entire process from scratch.

A single call to Claude has a cap on the total amount of content it can read in and write out, measured in tokens rather than word count — as length approaches that cap, a task doesn't fail loudly, it's more likely to get quietly truncated or have part of the content overlooked somewhere you didn't notice.

<a href="/en/glossary/core-concepts/xml-tag-structuring/">XML Tag Structuring</a> means wrapping different types of content in a prompt with tags like &lt;instructions&gt;&lt;/instructions&gt;, so Claude can clearly tell which part is background material, which part is a rule to follow, and which part is the actual task — instead of leaving everything mixed together in one <a href="https://crypto-bible.com/en/glossary/blockchain-fundamentals/block/" target="_blank" rel="noopener">Block</a> of text for Claude to guess at.

Skip giving any examples and rely entirely on the instruction itself to spell out the format, tone, and judgment criteria you want, so Claude produces the output directly — you save the time of preparing examples, but the instruction now has to carry the work examples would otherwise have done.