Reference · Updated weekly
The AI Coding Dictionary
Every word you'll hear around Claude Code, Codex, and Cursor — context windows, compaction, harnesses, handoffs — in plain English, with how each one shows up in the tools and where the school teaches it.
81terms
7sections
12school originals
Sep 12, 2026last updated
81 terms
§01
The Model
What a model is, what it isn't, and where the bill comes from.AIAn umbrella word that keeps changing what it points at. In coding today it means a language model plus the harness that lets it act.ModelThe trained network itself: billions of parameters that turn a context into the next token, and nothing more.Parametersaka WeightsThe billions of numbers a model is made of, fixed once training ends. Also called weights. What the model knows by heart lives in them.Trainingaka Pre-training, Post-trainingHow a model's weights get their values: predicting text at huge scale, then rounds of feedback that shape it into an assistant.InferenceUsing the model rather than training it. Every reply, edit and tool call you see comes from an inference pass; the weights never move.Effortaka Reasoning effort, Extended thinkingHow hard the model thinks before replying. Turn it up for tricky problems and pay in tokens and wait; turn it down for routine edits.TokenThe unit a model reads and writes in — a word fragment, a little shorter than a word — and the unit every limit and price uses.Next-token predictionThe model's one trick: pick a likely next token, stick it on the end, go again. Prose, code and tool calls all come out this way.Non-determinismaka NondeterminismRun the same prompt twice and you can get two different answers. Sampling, batching and silent provider updates all contribute.Model provideraka Provider, Inference providerThe service that actually runs the model: a lab like Anthropic or OpenAI, a cloud reselling it, or your own laptop through Ollama.Harnessaka Agent harness, ScaffoldThe software around the model that gives it hands: a system prompt, tools, permissions, hooks and the loop that feeds results back.Model provider requestaka API call, Model callA single API call: the harness ships the full context to the provider and gets one reply back. Most turns need several of them.Input tokensaka Prompt tokensEverything the model reads on a request: instructions, history, tool definitions, results. Cheap per token, but there are a lot of them.Output tokensaka Completion tokensWhat the model writes: replies, code, tool calls, hidden thinking. Produced one by one, so they set the wait, and they cost the most.Prefix cacheaka Prompt caching, Prompt cacheThe provider remembers the start of your last prompt, so the next request that begins the same way is faster and much cheaper.Cache tokensaka Cached tokens, Cache read tokensThe share of a request's input served from the prefix cache. Heavily discounted, and the first thing to check when a session feels pricey.Usage limitaka Rate limit, Usage capThe cap on how much agent work your plan or API key allows in a window. Hitting it pauses you; it isn't the context window.
§02
Sessions, Context Windows & Turns
The unit of work: what the agent sees, how long it lasts, and why it forgets.StatelessNothing carries over on its own: each model request starts from zero, and each new session does too.ContextEverything the agent currently has in front of it that bears on the task: loaded files, your messages, tool results so far.Context windowaka Context length, WindowThe fixed-size buffer of tokens a model can read in one request; if something isn't in it, the model can't see it.StatefulCarries information forward from one step to the next; a session is stateful across turns even though the model underneath is not.Agentaka Coding agent, AI agentA model wired to tools and a loop: it reads, acts, checks the result, and goes again until the task is done or you stop it.System promptaka System messageThe standing instructions the harness puts in front of every request: who the agent is, what tools it has, how it should behave.Sessionaka Conversation, ThreadOne continuous run with an agent, from an empty context to the moment you clear it, close it, or hand its work off.TurnYour message and all the work the agent does before it hands control back to you.Steeringaka Interrupting, RedirectingNudging a running agent back on course, mid-turn or between turns, before a wrong call becomes the foundation for everything after it.
§03
Tools & Environment
How an agent touches the world: files, shells, permissions, and sandboxes.EnvironmentEverything outside the harness the agent can inspect or change: your repo, your shell, the services it can reach, the browser it drives.Filesystemaka File system, Working directoryThe directory tree the agent reads, edits and runs inside; for a coding agent, the main part of the environment.Toolaka Function, Built-in toolA named capability the harness lets the model invoke, such as reading a file, running a shell command or fetching a page.Tool callaka Function call, Tool useThe model's request to run a tool: a structured message naming the tool and its arguments, which the harness then executes.Tool resultaka Tool outputWhat comes back from a tool call and lands in the context: file contents, command output, an error, a list of search hits.MCPaka Model Context Protocol, MCP serverModel Context Protocol: an open standard for plugging external tool servers into any agent harness.Permission requestaka Approval prompt, Permission promptThe harness pausing before a risky tool call to ask you yes or no; the simplest human-in-the-loop gate there is.Permission modeaka Auto-approve, Bypass permissionsThe setting that decides which tool calls run automatically and which stop for a permission request.Agent modeaka ModeA named preset that bundles a permission mode with behavioural instructions, switchable in the middle of a session.Plan modeaka Planning mode, Read-only modeA read-only agent mode: the agent may search and read, but must propose a plan for your approval before it edits anything.Sandboxaka Sandboxing, Isolated environmentAn isolated place for the agent to run, such as a container, VM or restricted shell, so a bad action can't reach the rest of your machine.Hooksaka Lifecycle hooks, Pre-tool hookCommands you configure the harness to run at fixed moments, such as before or after a tool call, regardless of what the model wants.Slash commandaka Custom command, Prompt shortcutA typed /name shortcut that runs a built-in action or expands a saved prompt, so the prompts you reuse live in the repo.Worktreeaka Git worktree, Parallel checkoutA second checkout of the same git repo in its own directory, so an agent can work on a branch without touching yours.Headless modeaka Non-interactive mode, Print modeRunning the agent from a script or CI with a prompt and no interactive UI; the result comes back as text or JSON.Checkpointaka Rewind, SnapshotA saved state of your files, and sometimes the conversation, that you can rewind to after a turn goes wrong.
§04
Failure Modes
The ways it goes wrong: confident nonsense, stale knowledge, and long-session drift.Sycophancyaka Yes-man behaviour, AgreeablenessThe model's tilt toward agreeing with you, praising your plan, and telling you it worked, regardless of whether it did.Hallucinationaka Confabulation, FabricationOutput that is fluent, confident, and wrong: an invented API, a misquoted file, a test result that never happened.Parametric knowledgeaka Baked-in knowledge, Training knowledgeWhat the model knows because it was in the training data, stored in its weights and frozen from that moment on.Knowledge cutoffaka Training cutoff, Cutoff dateThe date the model's training data ends; anything released after it is unknown to the model unless you load it into context.Contextual knowledgeaka In-context knowledgeFacts the model has because they are in the context window right now, as opposed to facts it remembers from training.Attention relationshipThe link between any two tokens in the context; the model weighs each pair, and there are far more pairs than tokens.Attention budgetThe fixed amount of focus each token can spread across the rest of the context; more context means thinner slices.Attention degradationaka Context rot, Lost in the middleThe slow drop in output quality as a session grows and every token's attention is spread across more competing material.Smart zoneaka Dumb zone, Smart zone / dumb zoneThe early stretch of a session where the agent is at its sharpest; past it, the same model gets sloppier and forgetful.Prompt injectionaka Indirect prompt injection, Tool-result injectionInstructions smuggled into something the agent reads, which the model may follow as if they came from you.
§05
Handoffs
Moving work between sessions without losing the plot.Clearingaka /clear, Fresh sessionWiping the session so the next request starts from an empty context window, keeping only the standing instructions.Handoffaka Hand-off, Session handoffEnding one session and starting another on the same task, carrying the state forward in writing rather than in the window.Primary sourceThe real thing: the file, the diff, the test output, the docs page. What the agent should read instead of a description of it.Secondary sourceAn account of the thing rather than the thing itself: a summary, a paraphrase, a compaction. Useful, but one step removed from the truth.Handoff artifactaka Handoff doc, Handoff noteThe written note that carries a task across a session boundary: what's done, what's left, what was decided, where to look.Specaka Specification, PRDA written description of what to build and how you'll know it's right, agreed before the agent starts writing code.Ticketaka Issue, TaskA bounded unit of work with a clear done state, sized so an agent can finish it in one session.Compactionaka /compact, Context summarizationReplacing the session so far with a shorter summary of it, freeing the context window at the cost of detail.Autocompactaka Auto-compaction, Automatic compactionCompaction the harness triggers on its own when the context window nears its limit, without asking you first.
§06
Memory & Steering
Getting instructions to stick across sessions, and pointing the agent at the right thing.Memory systemaka Memory, Persistent memoryWhatever lets an agent carry information across sessions: notes on disk, a memory directory, a store it reads at startup.AGENTS.mdaka CLAUDE.md, Rules fileThe instructions file at the repo root that every session reads first: commands, conventions, gotchas. CLAUDE.md in Claude Code.Progressive disclosureGiving the agent a short index up front and letting it load the detail only when a task calls for it.Context pointerA short reference that tells the agent where information lives and when to fetch it, instead of the information itself.Context engineeringaka Context management, Context curationDeciding what goes into the context window, when, and in what order, so the agent sees what the task needs and nothing else.Skillaka Agent skill, SKILL.mdA named, reusable procedure the agent loads on demand: a folder of instructions and optional scripts, used when a task matches.Subagentaka Sub-agent, Agent toolA separate agent the main agent spins up for a bounded task, with its own context window; only its final report comes back.
§07
Patterns of Work
The working habits that separate a good run from a bad one.Human-in-the-loopaka HITL, Supervised runA working pattern where a person approves, corrects, or answers for the agent while it runs, instead of only judging the result.AFKaka Away from keyboard, Unattended runAway from keyboard: you start the agent, leave, and come back to finished work you review later instead of supervising live.Automated checkaka Check, CI checkA mechanical pass/fail test the harness or CI runs on the agent's work: types, lint, tests, build. Cheap, fast, no judgement.Automated reviewaka AI code review, Agent reviewA model reads the agent's diff and flags problems before a person does: judgement without a human, fuzzier than a check.Human reviewaka Code review, Diff reviewA person reading the agent's work before it ships. The only step that judges whether the change is right, not just whether it passes.Vibe codingaka VibecodingBuilding software by describing what you want to an agent and judging the result rather than reading every line of code.One-shotaka One-shotting, One-shot promptGetting a usable result from a single prompt with no follow-up turns. A good test of a prompt, a bad habit for production.Design conceptaka Design brief, Concept docA short written description of what you're building and why, agreed before any code, so the agent and you share the same picture.Grillingaka Grill me, InterrogationHaving the agent interview you, question by question, until the requirements are fully resolved before it writes anything.Prototypingaka Throwaway prototype, SpikeBuilding a quick, throwaway version to learn something, not to keep. With agents it's cheap enough to do before deciding anything.Harness engineeringaka Environment engineering, Agent opsImproving the setup around the model (tools, checks, instructions, permissions) so the same model does better work.DXaka Developer experienceDeveloper experience: how pleasant and fast a tool, codebase or workflow is for the humans using it.AXaka Agent experienceAgent experience: how well a codebase, tool or environment supports an agent that starts each session knowing nothing.
Recently added
Changelog
- Dictionary launched: 81 entries across 7 sections, with tool-specific notes and course links.
How this dictionary works
Seven sections, read in order, take you from what a model is to how to run agents without watching them. Every definition is written here in the school's voice; the term list tracks the open Dictionary of AI Coding plus the words we use in class, and a scheduled agent re-checks both every week. Looking for transformers, LoRA, or embeddings instead? That's the AI Glossary. Missing a word? Open an issue.
Words are the easy part
Course 01 puts every one of these terms to work: you install Claude Code, run the loop, and ship a real project — permission modes, compaction, hooks and all.
Start Course 01 →