AI Vocabulary | Work Is Code

Last updated: April 28, 2026.

A reference for the terms that come up across these articles. Some are foundational AI and developer jargon. Others are concepts specific to this site that have their own dedicated articles, in which case the entry summarizes and points there. If you want a precise read on something you saw in a piece, this is where to come.

The stack

Term	Definition
LLM (Large Language Model)	An LLM is a neural network trained on large volumes of text to predict and generate language. It is the category of AI that powers Claude, ChatGPT, Gemini, and other modern AI assistants.
Model	The model is the actual AI (Opus, Sonnet, or Haiku for Claude). It runs inside a harness and uses the tools the harness provides. The harness is the body. The model is the brain.
Harness	The harness is the runtime that runs the model and exposes tools to it. Claude Code is a harness whether it is running as a CLI, an IDE extension, or via the SDK. Cursor, Cline, and Aider are other harnesses.
Tool	A tool is a function the harness exposes to the model, such as reading a file, running a bash command, searching the web, or editing code. Tools are how the model actually does things in your environment.
Token	A token is the unit the model thinks in, roughly equivalent to 0.75 of an English word. Tokens are also the unit AI providers use for API billing, priced per million for input and output separately. Long sessions, big files, and verbose responses fill the context window and increase API cost in lockstep.
Context window	The context window is the model’s working memory for the current session, measured in tokens. It is finite, so once you fill it, performance degrades. Subagents and `/compact` are how you protect it.

Editors and runtimes

Term	Definition
IDE (Integrated Development Environment)	An IDE is a code editor with built-in tools, including a file tree, debugger, terminal, and extensions. VSCode, the JetBrains family (IntelliJ, PyCharm, WebStorm), and Cursor are IDEs. Claude Code can run as an extension inside one.
CLI (Command-Line Interface)	A CLI is a text-based way to drive software from a terminal. Claude Code’s CLI is the `claude` command. It is useful for terminal-first workflows, scripting, and CI.
SDK (Software Development Kit)	An SDK is a library you import into your own code to use a service programmatically. The Claude Agent SDK lets you build your own harness on the same engine that powers Claude Code.

Extending the harness

Term	Definition
MCP (Model Context Protocol)	MCP is a standard for connecting external services to the harness as tools. Slack, calendar, databases, and browsers can all be MCPs. Each runs as a separate background process, so RAM cost is real.
Skill (slash command)	A skill is a reusable workflow defined as a markdown file with a prompt and a few rules. You invoke it with `/name`. Some skills ship with the harness, and you can write your own.
Hook	A hook is a shell command the harness runs automatically at lifecycle events: before a tool runs, after a tool runs, on session start, or on stop. Hooks enforce behavior at the harness level instead of relying on the model to remember.
Plugin	A plugin is a bundle of skills, hooks, and commands shipped together as one installable unit. Plugins are how power-user setups travel.
Subagent	A subagent is a sub-process the main agent spawns for one task. It has its own context window and returns just the answer to the parent. Use one when you want to research something without cluttering your main conversation.

Persistence and configuration

Term	Definition
CLAUDE.md / AGENTS.md	A CLAUDE.md (or AGENTS.md) is a markdown file at the repo root that the harness loads at session start. It tells the model who you are, the project context, and your hard rules. Different harnesses use different filenames (`CLAUDE.md`, `AGENTS.md`, `.cursorrules`), but the idea is the same.
STATE.md	STATE.md is a living file at the repo root that records current project status: active blockers, in-flight decisions, what changed last session, and what needs to happen next. It is the snapshot that lets any agent or collaborator walk into the project mid-stream without a briefing.
Memory	Memory is a persistent personal store outside any single session. It auto-loads at session start. It differs from CLAUDE.md, which is repo-level and committed; memory is personal and follows you across every session in a project.
Permissions	Permissions is a `settings.json` structure that pre-approves (or denies) tools and commands so you stop getting prompted on every routine action. It lives in `.claude/settings.json` (shared) or `.claude/settings.local.json` (personal).
Plan mode	Plan mode is a setting where the model writes a plan before it can edit anything. It forces planning instead of one-shot editing. You toggle it with shift+tab in Claude Code.
ADR (Architecture Decision Record)	An ADR is a short document that records a significant architectural decision: what was decided, why, and what alternatives were rejected. ADRs are dated, immutable, and superseded rather than deleted when decisions change.
Runbook	A runbook is a documented procedure for a recurring operational task, such as deploying a release or responding to an incident. It encodes the steps so the person executing them does not need to derive them from first principles each time.

Failure modes

Term	Definition
Hallucination	Hallucination is the failure mode where a model generates confident, fluent output that is factually wrong or entirely fabricated, without signaling uncertainty. It occurs because the model predicts plausible text rather than retrieving verified facts.
Sycophancy	Sycophancy is the tendency of models to optimize for user approval rather than accuracy, producing agreeable, validating responses even when the correct answer would contradict the user. It is a structural consequence of training on human feedback rather than a deliberate design choice.
Persuasion bombing	Persuasion bombing is the pattern where a model, when challenged, escalates rhetorical intensity instead of genuinely reconsidering: adding headers, bullets, flattery, and unsolicited data to defend its original position. See Persuasion Bombing for the cross-model field test.
Automation bias	Automation bias is the documented tendency for people to over-accept computer output as a cognitive shortcut, especially under high workload, time pressure, and low confidence. It is the mechanism that converts AI overwhelm from an annoyance into a systemic decision-quality risk.

Disciplines and antipatterns

Term	Definition
Specification	Specification is the act of saying what you want with enough precision that the machine can act on it. It is the non-coding half of “the job has changed.”
Validation	Validation is the act of knowing whether what came back is right. It is the other non-coding half.
Vibe coding	Vibe coding is Andrej Karpathy’s term for AI-driven coding without specification or validation discipline. It works fast for prototypes but fails on anything that needs to keep working. Vibe coding is the opposite of AI-native engineering.
Vibe engineering	Vibe engineering is the intentional counterpart to vibe coding. Architecture governs AI execution, every session produces working software and documentation, and knowledge compounds across sessions rather than evaporating. See AI-Native Engineering for the full doctrine.
Project Brain	The Project Brain is the persistent layered cognition system (CLAUDE.md, STATE.md, ADRs, runbooks, wiki) that gives an AI the same context a senior engineer would have after six months on a project. Without it, AI-accelerated development degrades into vibe coding.
Technical debt	Technical debt is the accumulated cost of shortcuts, deferred fixes, and architectural corners cut during development. It compounds over time, slowing velocity and increasing the cost of every subsequent change until it is explicitly paid down.
Cognitive debt	Cognitive debt is the growing gap between what you have built and what you understand about what you have built. It accumulates when AI-accelerated output velocity exceeds your capacity to capture decisions, patterns, and reasoning across sessions. See The Debt You Don’t See for the full argument.
AI-accelerated Kanban	AI-accelerated Kanban is a session-based process model used in place of sprint-based iteration. The premise: when AI compresses build time to a fraction of its prior cost, two-week sprints become waiting rooms rather than planning tools. See AI-Native Engineering.

Frameworks specific to this site

These have their own articles. The entries below are summaries; follow the links for the full treatment.

Term	Definition
Translation layer	A translation layer is a system or person that converts organizational knowledge from its native format into whatever format the audience in front of it can actually use. The argument is that AI is becoming this layer for entire organizations, eliminating the manual retyping that currently sits at every format boundary. See The AI Translation Layer.
Knowledge lock	The Knowledge Lock is the invisible permissions system every organization runs by accident, enforced by format rather than policy. When information exists only in a format an audience cannot open, the knowledge is effectively inaccessible even though no one explicitly denied access. See The Knowledge Lock.
Story point pricing	A story point is a unit of estimated complexity used in agile development, not tied to hours. Story points are the pricing unit for AI-native work on this site because they measure the complexity of what was solved rather than the time it took. See Story Point Pricing.