Glossary
The vocabulary that comes up when directing a model to build something real, from the first prompt to the live URL. Start with Vibe Coding 101 for the walkthrough these terms describe.
A
·2 TERMSagent /noun/ — a model that plans its own next steps and calls tools to get there, instead of answering in a single turn. Claude Code is the agent behind every build documented on this site.
authority level /noun/ — a recipe's authorityLevel frontmatter field (primary or secondary) — exactly one primary recipe per pillar/topic, the rest secondary. Surfaces as an authority-level meta tag for crawlers and, on the page itself, as a plain-text field in the metadata toggle.
B
·1 TERMbuild /noun, verb/ — compiling the actual project from source (npm run build here). A clean build is a correctness gate to clear before a pull request opens, not a nice-to-have.
C
·4 TERMScommit / pull request (PR) /noun/ — a commit is one saved, reviewable unit of change; a PR is the request to merge it into the main branch, and the point where a build either passes or it doesn't.
context /noun/ — everything a model can see when it answers: open files, the conversation so far, project instructions. More context isn't automatically better context — irrelevant context crowds out what actually matters.
context window /noun/ — the hard ceiling on how much context fits at once, measured in tokens. Long sessions get summarized once they approach it, which is why durable decisions belong in the repo, not the conversation.
cosine similarity /noun/ — a score from -1 to 1 measuring how closely two embedding vectors point in the same direction, ignoring their length: 1 means near-identical meaning, 0 means unrelated. The actual math behind semantic search once two pieces of text have been turned into vectors.
D
·2 TERMSdeploy /verb/ — publishing a build somewhere reachable. On this site, every push triggers its own Cloudflare build before anything reaches the live domain.
DOM (Document Object Model) /noun/ — the browser's in-memory tree of a page's elements, built from the HTML source and updated live as scripts run. What a crawler sees by rendering the DOM can differ from the raw HTML response — relevant for content that JavaScript adds after load.
E
·1 TERMembedding /noun/ — a vector of numbers (768 of them, in this site's own semantic search layer) that a model produces from a piece of text, positioning it in a shared space where nearby vectors mean similar meaning. Two pages about the same topic land close together even if they never use the same words.
F
·1 TERMfrontmatter /noun/ — the block of structured metadata at the top of a file, set off by --- fences and usually written in YAML — a recipe’s title/result/tools, a Claude Skill’s name and trigger description, this site’s own CLAUDE.md. Read before the body, not part of what a reader sees.
H
·1 TERMhallucination /noun/ — a confident, wrong answer with no visible sign that it’s wrong. The reason “does it build” and “does the number check out” matter more than “does it sound right.”
J
·1 TERMJSON-LD /noun/ — structured data embedded in a page as a <script type="application/ld+json"> block, describing what the page is (an article, a FAQ, an organization) in a format search engines and AI crawlers can parse directly instead of inferring it from prose.
L
·1 TERMlint /noun, verb/ — a deterministic check that flags code or copy issues by pattern- or rule-matching, without calling a model. Cheaper and more consistent than a model call for anything with a fixed, checkable answer — like a forbidden-word list.
M
·3 TERMSmeta description /noun/ — the <meta name="description"> tag on a page — a one- or two-sentence summary that search engines often show as the snippet under a search result, and that this site also reuses for the OpenGraph and Twitter card descriptions.
MCP (Model Context Protocol) /noun/ — an open standard that lets an agent call out to external tools and data sources through one common interface, instead of every integration being bespoke.
MCP server /noun/ — a standalone process built on MCP that exposes resources, prompts, or tools over a common interface, so any compatible client — Claude Desktop, a different agent, a CI job — can call it without cloning the repo it came from.
N
·1 TERMnpm /noun/ — Node's default package manager — installs the dependencies in package.json and runs the scripts defined there, like npm run build, the one this site treats as its correctness gate before a PR opens.
O
·1 TERMOpenGraph /noun/ — the og:* meta tags (title, description, type, image) that control how a link looks when it’s pasted into Slack, iMessage, or a social feed, instead of falling back to a bare URL.
P
·4 TERMSpattern /noun/ — the reusable technique behind a recipe: the part you can actually take and reapply, independent of the specific tool that shipped it.
pgvector /noun/ — a PostgreSQL extension that adds a vector column type and nearest-neighbor search (cosine, L2, inner product) directly to a Postgres database, so cosine similarity over embeddings runs as a real SQL query with an index instead of a scan over a flat file. The natural next step once a semantic search outgrows a JSON file a few dozen rows long.
prompt /noun/ — the instruction actually sent to the model. The single variable most worth rewriting before touching any code.
provenance /noun/ — a record of who authored a given block and how (human, Claude, or a mix of both), plus the prompt that could rebuild it. This site’s own version sits behind the “Behind the build” button on every page.
R
·2 TERMSrecipe /noun/ — this site's unit of content: a pattern paired with a documented, real result, not a general tips post.
repo (repository) /noun/ — where code and its history actually live. The shared state that survives after any one agent's session ends.
S
·4 TERMSsemantic search /noun/ — finding results by meaning instead of exact keyword match: text gets converted into an embedding (a vector of numbers), and two pieces of text count as similar when their vectors point in a close direction, measured with cosine similarity. This site's own semantic layer (src/data/semantic-layer.json) embeds every public page this way, so related content turns up even when the pages don't share any of the same words.
SEO metadata /noun/ — the non-visible tags in a page's <head> — meta description, OpenGraph, JSON-LD, canonical URL — that tell search engines and AI crawlers what a page is about, separate from what a human reader actually sees.
slop (AI slop) /noun/ — copy or design that reads unmistakably AI-generated: left-accent boxes, an eyebrow over everything, an icon on every bullet. Cataloged in detail on Vibe Check.
system prompt / project instructions /noun/ — the standing instructions a model reads before anything else in a session (this repo's own CLAUDE.md). Durable across sessions, unlike a single prompt.
T
·2 TERMStoken /noun/ — a model's unit of text, roughly a word or a word-piece. Usage costs and context-window limits are both counted in tokens, not words.
tool call /noun/ — the moment an agent stops writing text and does something with real effects instead: running a command, editing a file, calling an API.
V
·1 TERMvibe coding /noun/ — describing what you want in plain language and letting a model do the typing. The result depends less on how the prompt is worded than on how much real context it’s given — see Vibe Coding 101.
Y
·1 TERMYAML /noun/ — a plain-text, indentation-based format for structured data — the frontmatter block at the top of every recipe in this site’s src/content/recipes/ (title, result, tools) before the body starts.