Dictionary

Plain definitions for context engineering and token economics.

Start here

TokenA small chunk of text, often part of a word, that AI models read, count, and charge by.Context WindowThe amount of text an AI can keep in mind at once, like the size of a desk where everything must fit.Prompt CachingThe AI service remembers the unchanged start of your request so repeat requests are cheaper and faster, like a saved draft.AgentAn AI that does not just answer but takes actions, checks the results, and keeps going until the job is done.Model Context Protocol (MCP)A universal plug standard that lets AI apps connect to tools and data, like USB-C for AI.BenchmarkA standard exam for AI models: the same questions for every model, graded automatically, so you can compare scores fairly.HallucinationWhen an AI confidently makes something up, like a fake fact, a library that does not exist, or a wrong quote.CompactionShrinking a long conversation into a short summary so the AI can keep working, like condensing meeting notes.

A

Adaptive ThinkingToken EconomicsThe AI decides for itself how long to think about each request, guided by how much effort you ask for.

AgentAgentsAn AI that does not just answer but takes actions, checks the results, and keeps going until the job is done.

Agent HooksAgentsScripts that always run at set moments in an AI session, like a doorbell that rings every time someone enters.

Agent OrchestrationAgentsSeveral specialist AIs working together under a coordinator, like a project manager directing a small team.

Agent SkillsAgentsPackaged know-how the AI loads only when a task needs it, like a recipe card pulled from a box.

Agentic LoopAgentsThe repeat cycle of think, act, check the result, and think again that an AI agent runs until it finishes.

AGENTS.mdContext ManagementAn instruction sheet in a code project written for AI helpers, like a README but for agents.

API-Equivalent CostToken EconomicsWhat your AI usage would have cost at pay-per-token prices, useful even when you pay a flat subscription.

ARC-AGIEvaluationPicture puzzles easy for people but hard for AI, testing whether it can learn a new rule from a few examples.

Attention BudgetContext ManagementHow much the AI can truly focus on, which is often less than how much it can technically hold.

Auto MemoryMemory StrategiesThe AI notices useful facts while working and saves them for next time, without you writing anything down.

Automatic Context CompactionMemory StrategiesThe AI service itself summarizes old parts of a long chat for you, so the conversation never runs out of room.

B

Batch ProcessingToken EconomicsSending many requests to be processed later, in bulk, at a lower price, like standard shipping versus express.

BenchmarkEvaluationA standard exam for AI models: the same questions for every model, graded automatically, so you can compare scores fairly.

Benchmark ContaminationEvaluationWhen an AI has already seen the test answers during training, so its high score says less about real ability.

Benchmark SaturationEvaluationWhen an exam becomes so easy for top AIs that everyone scores near 100% and it stops ranking them.

Break-Even PointCachingHow many times you must reuse stored text before caching saves more than it costs, like a gym membership.

C

Cache Invalidation HierarchyCachingChange something early in a request and everything after it must be cached again, like editing page one of a book.

Cache ReadCachingThe discounted price for reusing text the AI service already stored, often a tenth of the normal price or less.

Cache TTLCachingHow long stored text stays ready for cheap reuse before it expires, like a parking meter.

Cache WriteCachingThe one-time extra fee to store part of your request for reuse, like paying to rent a locker.

Chain of Thought (CoT)Prompt EngineeringAsking the AI to think step by step before answering, like showing your working in a math exam.

CLAUDE.mdContext ManagementA notes file Claude Code reads at the start of every session, so you never repeat the basics.

Cobb-Douglas ModelEvaluationAn economics formula used to estimate how much answer quality you get for the tokens you pay for.

Code MapContext ManagementA table of contents for a codebase, so the AI knows where things live before it opens any files.

CompactionContext ManagementShrinking a long conversation into a short summary so the AI can keep working, like condensing meeting notes.

Compaction TriggerMemory StrategiesThe fullness level at which automatic summarizing kicks in, like a fuel gauge warning light.

Context EditingMemory StrategiesAutomatic rules that tidy up old clutter in a conversation before the AI reads it, with no extra AI involved.

Context EngineeringContext ManagementChoosing exactly what information to hand the AI, like packing only what you need for a trip.

Context PollutionContext ManagementClutter in the AI's working memory that makes its answers worse, like a messy desk hiding the paper you need.

Context RotContext ManagementThe longer a conversation gets, the worse the AI recalls details from early on, like a fading memory.

Context WindowContext ManagementThe amount of text an AI can keep in mind at once, like the size of a desk where everything must fit.

Copilot MemoryMemory StrategiesGitHub Copilot jots down facts about your repository and preferences, and forgets ones it has not used in a while.

Cost per Million Tokens (MTok)Token EconomicsThe standard price tag for AI usage: how much one million tokens of text costs to send or receive.

Custom AgentAgentsA ready-made AI persona with its own instructions and tools, picked from a menu, like choosing a specialist.

Custom Instructions FileContext ManagementA file in your repository that GitHub Copilot automatically adds to every chat, like house rules on the fridge.

D

DistractorsContext ManagementInformation that looks relevant but is not, like a misleading road sign that sends the AI the wrong way.

E

EffortToken EconomicsA dial that sets how hard the AI works on a task, trading thoroughness for speed and cost.

EmbeddingArchitectureTurning text into a list of numbers that captures its meaning, so a computer can find similar ideas, not just matching words.

Extended ThinkingPrompt EngineeringLetting the AI quietly reason before it replies, the older way of switching this on with a fixed thinking allowance.

F

Few-Shot PromptingPrompt EngineeringShowing the AI a few examples of what you want before asking, like showing a sample letter before requesting one.

Fine-TuningArchitectureGiving an AI extra training on your own examples so it permanently picks up a style or specialty.

Function Calling / Tool UseArchitectureLetting the AI ask your software to do things, like search, read a file, or run a command, and use the results.

G

GPQAEvaluationVery hard science questions that even smart people with Google struggle to answer, used to test expert-level AI knowledge.

H

HallucinationEvaluationWhen an AI confidently makes something up, like a fake fact, a library that does not exist, or a wrong quote.

High-Signal TokensToken EconomicsGiving the AI only the information that truly helps, like a short, precise brief instead of a thick folder.

Humanity's Last ExamEvaluationA very hard exam of 2,500 expert questions across many subjects, built because AIs were acing older tests.

I

Input TokensToken EconomicsEverything you send to the AI, including instructions, history, and files, counted in tokens.

J

JIT Context (Just-in-Time)Context ManagementFetching information at the moment it is needed rather than loading everything up front, like cooking to order.

K

Knowledge CutoffContext ManagementThe date when an AI's learning stopped; it does not know about anything newer unless you tell it.

L

Lightweight IdentifiersContext ManagementKeeping a note of where things are, like a library card number, and fetching the book only when needed.

LLM-as-a-JudgeEvaluationUsing one AI to grade another AI's answers, like a teacher marking essays, which is fast but has its own biases.

Long-Context PricingToken EconomicsSome models charge a higher rate for the whole request once your prompt passes a size limit, like a surcharge.

Lost in the MiddleContext ManagementAIs pay most attention to the start and end of a long text and tend to miss what is in the middle.

M

Max Output TokensToken EconomicsA cap on how long the AI's reply can be; if it hits the cap, the answer simply stops.

MCP AppsArchitectureSmall interactive screens, like charts or forms, that a connected tool can show right inside the AI chat.

Memory ToolMemory StrategiesA filing cabinet the AI can write notes into and read from later, so it remembers things between sessions.

Message SteeringAgentsSending a quick correction while the AI is busy, so it changes course without starting over.

Mid-Conversation System MessagePrompt EngineeringAdding a new instruction partway through a chat without rewriting the original setup, so cached savings stay intact.

Minimum Cacheable PrefixCachingText must be at least a certain length before it can be cached; shorter text is simply processed normally.

Model Context Protocol (MCP)ArchitectureA universal plug standard that lets AI apps connect to tools and data, like USB-C for AI.

O

Observation MaskingMemory StrategiesReplacing old tool outputs with short placeholders instead of summarizing them, a cheap way to free up space.

Output TokensToken EconomicsThe words the AI writes back to you, which usually cost several times more than what you send.

P

pass@kEvaluationThe chance an AI solves a task in at least one of k tries, like keeping a student's best attempt.

Path-Scoped RulesContext ManagementInstructions that only apply to certain folders or file types, like rules posted on one room's door.

PluginArchitectureOne installable package that adds commands, skills, and helpers to Claude Code at once, like an app bundle.

Preserved ThinkingArchitectureThe AI's earlier reasoning is locked to the exact conversation that produced it, so rewriting history breaks it.

Programmatic Tool CallingArchitectureThe AI writes a small program that calls tools for it, so only the final answer comes back to read.

Progressive DisclosureContext ManagementRevealing information to the AI only when it needs it, like turning pages instead of dumping the whole book at once.

PromptPrompt EngineeringEverything you give the AI in one go: your question plus the instructions and background that come with it.

Prompt CachingCachingThe AI service remembers the unchanged start of your request so repeat requests are cheaper and faster, like a saved draft.

Prompt FilePrompt EngineeringA saved, reusable request you can run with a short command instead of retyping it each time.

Prompt InjectionArchitectureHidden instructions slipped into text an AI reads, tricking it into doing something its user never asked for.

R

RAG (Retrieval-Augmented Generation)ArchitectureLooking up relevant documents and handing them to the AI before it answers, like an open-book exam.

Rate LimitToken EconomicsA speed limit on how many requests or tokens you can send per minute; go over it and you must wait.

S

Session TranscriptMemory StrategiesA complete log of everything said and done in an AI coding session, kept as a file on your computer.

Structural PatternsContext ManagementHow you arrange information matters: the AI does better with tidy, logical ordering than with a jumbled pile.

Structured Note-takingContext ManagementThe AI keeps a running to-do list in a file, like a sticky note it can reread after a break.

Structured OutputsArchitectureForcing the AI to answer in an exact data format, like filling in a form, so programs can read it reliably.

SubagentAgentsA helper AI the main AI sends off to do one job, which reports back only a short result.

SummarisationMemory StrategiesReplacing the full conversation with a short recap before each new message, like reading the minutes instead of the transcript.

Summary PromptMemory StrategiesYour own instructions for what a summary must keep, like telling a note-taker which details matter most.

SWE-benchEvaluationA test where AI fixes real bugs from open-source projects, checked by running each project's own tests.

System PromptPrompt EngineeringStanding instructions given to the AI before the conversation starts, like a job description it follows throughout.

T

Tab Relevance ScoringArchitectureRanking your open files by how useful they are for the current task, so the least useful can be closed.

Task BudgetToken EconomicsA token allowance the AI can see counting down, so it paces itself and wraps up before running out.

TemperaturePrompt EngineeringA creativity dial: low makes the AI's answers predictable and focused, high makes them more varied and surprising.

Terminal SandboxingArchitectureA fenced-off area where the AI's commands run, so they cannot touch files or websites outside it.

Terminal-BenchEvaluationA test of whether an AI agent can get real jobs done in a computer terminal, like a sysadmin practical exam.

Thinking TokensToken EconomicsThe hidden scratch work the AI does before answering, which you pay for like regular output.

Time to First Token (TTFT)EvaluationHow long you wait before the AI starts replying, like the pause before someone begins to answer.

TokenToken EconomicsA small chunk of text, often part of a word, that AI models read, count, and charge by.

Token Budget DashboardArchitectureA live gauge showing how full the AI's memory is and what is filling it, like a phone storage screen.

Tokenizer DriftToken EconomicsDifferent model versions can count the same text as a different number of tokens, which changes the bill.

Tool Result ClearingContext ManagementThrowing away bulky old printouts while keeping the conclusions you wrote down from them.

Tool SearchArchitectureInstead of describing every tool up front, the AI searches a catalog and loads only the tools it needs.

Trimming (Last-N)Memory StrategiesKeeping only the most recent messages and forgetting older ones, like a chat that shows just the last few texts.

Turn AnalyticsArchitectureA forecast of how many more back-and-forth messages fit before the AI's memory fills up.

X

XML TaggingPrompt EngineeringWrapping parts of your instructions in labeled tags so the AI can tell them apart, like folders with name tabs.