Glossary

Vocabulary you'll need.

76 terms. Linked from inside the chapters. Search Cmd-K from anywhere.

Jump by letter — sidebar tracks where you are.

Jump to section tap to open

Abliteration: A technique (coined by FailSpy in 2024, automated by the Heretic CLI in 2025) for permanently removing a model's refusal behavior by orthogonalizing every weight matrix that writes into the residual stream against the empirically-found refusal direction. Produces a "heretic model" — same weights minus one specific behavior. See The heretic question.
AEO: Answer Engine Optimization — structuring content to be the direct answer a search or AI surface returns (featured snippets, FAQ, 40–60-word answers), not just a blue link. Sibling of GEO.
Agent: An LLM in a loop with tools, working toward a goal across multiple turns. The bar isn't "uses tools once" — it's "decides for itself what to do next."
Agentic SEO: Using AI agents as the labor to build and maintain an SEO footprint end to end — research, pages, internal links, monitoring. The agents are cheap; the moat is a per-page differentiated dataset, not the fact that an agent wrote it. Google judges the output, not the production method.
Agent memory: What an agent carries between sessions. In Claude Code it is four layers — CLAUDE.md, memory/, skills, and the session itself. The hard part is not writing it; it is curating it so it stays deduped and true.
Agent teams: The other shape of Claude Code parallelism: a roster you define up front (one lead session plus named roles like Frontend Specialist, Backend Engineer) — for work that decomposes cleanly into domains. Contrast dynamic workflows, where Claude writes the decomposition itself.
Auto mode: No prompts, no approvals — --dangerously-skip-permissions (nuclear), an --allowed-tools allow-list (what pros use), or --auto (a classifier). The flag isn't the problem; the environment is.
Batch API: Half-price async inference if you can wait (up to 24h). The cost lever for non-interactive bulk work.
Cache breakpoint: The cache_control point in a prompt marking the cached prefix. Put volatile content (new CLAUDE.md sections) behind it or you pay the write premium for a discount you never collect.
Claude Code: Anthropic's CLI tool. `claude` in your terminal. Reads/writes your repo, spawns subagents, talks to MCP servers.
CLAUDE.md: A markdown file loaded on every turn of every session. Stacks: global ~/.claude/CLAUDE.md + repo-local + subdirectory, nearest wins. Under 100 lines — every line is billed in perpetuity and edits void the prompt cache.
Connector: Friendly UI for an MCP server. In Cowork & the Claude apps, "connector" is what you see; MCP is what's underneath.
Context window: The total tokens an instance can hold at once. The bigger the window, the more you can stuff in — but more isn't free.
Contrast ratio: The measured luminance ratio between text and its background, from 1:1 to 21:1. WCAG AA wants at least 4.5:1 for body text — a number, not an opinion, which is why a check catches what the eye approves.
Cowork: Anthropic's desktop app for knowledge work. Connectors, skills, scheduled tasks, mounted folders.
Cron: A scheduling syntax. `0 7 * * 1-5` = 7 AM weekdays. Most modern surfaces hide it behind a UI.
Design system: The written-down rules a product's look obeys — color/spacing/type tokens, component recipes, and the principles behind them. Your past taste decisions encoded so the next screen, or the next model, starts from your judgment instead of the internet's average.
Design token: A named design value — an accent color, a spacing step, a radius — stored once and referenced everywhere, so a theme change is one edit. The atom of a design system; the W3C DTCG format is the portable spec.
Dreaming: An automated memory-curation pass — an agent reads its own recent sessions, finds patterns, and improves its persistent memory. Anthropic ships it in Managed Agents (a research preview as of mid-2026), where it can auto-update memory; the local Claude Code build in Ch 44 is deliberately propose-only — it surfaces candidates and never writes.
Dynamic workflows: Opus 4.8's headline Claude Code feature: Claude writes a JavaScript script that plans a big task, fans out subagents (up to 16 at once, 1,000 per run), runs an adversarial verifier pass, and returns one answer — the orchestration runs in the background, outside your context. A research preview. See Dynamic Workflows.
/effort: The Opus 4.8 slash command that sets how hard the model thinks before it acts: low → medium → high (the default) → xhigh → max, with a separate ultracode notch. Most people leave it on the default and never feel the ceiling. See Dynamic Workflows — turning it on.
Eval: A test for your AI workflow. If you don't have evals, you don't have a workflow — you have a hope.
Evaluator: The small fast model (Haiku by default) that judges a /goal condition from the transcript after each turn. It has no tools — if your output can't prove "done," it loops forever.
Fable 5: Anthropic's most capable generally available model (released June 9, 2026; model id claude-fable-5). A Mythos-class model made safe for general use: safety classifiers gate offensive-cyber, bio/chem, and distillation requests — fallback to Opus 4.8 is built into the Claude apps and Managed Agents, opt-in on the API (blocked API requests error, unbilled). $10/$50 per Mtok, 1M context, 128K output. Also available as an advisor model cheaper workers call mid-task. See the model file.
Fork: Resuming an old session and submitting a new prompt grows a new branch — the original timeline survives on disk. You can't overwrite; every new prompt on an old session is a new branch.
Function calling: A model output format that says "I want to call X with Y." Your code does the actual call.
GEO: Generative Engine Optimization — earning a citation inside an AI-generated answer (ChatGPT, AI Overviews, Perplexity). Driven by evidential authority — stats, named sources, quotable claims — not keyword density. The one term in this family with a peer-reviewed origin.
/goal: /goal <condition> runs turns until a small evaluator reading the transcript says the condition holds. Removes per-turn approval the way Auto mode removes per-tool. Always add an "or stop after N turns" clause.
Hallucination: When the model confidently invents something. Less common in 2026 than 2023, still possible. Verify destructive actions.
Headless mode: Running Claude Code with no interactive UI via claude --print (-p). Prints to stdout, exits — pipeable, cron-able, drops straight into a GitHub Action.
Heretic model: An open-weights model whose refusal direction has been surgically removed via abliteration — orthogonalizing weight matrices against the linear refusal direction in the residual stream (Arditi et al., 2024). Not jailbreaking: the weights themselves no longer encode the refusal, persistent across all sessions. Trade-off: 1–6% benchmark regression, and the model is no longer aligned by its lab — alignment becomes your problem. See The heretic question.
Hook: A command, HTTP endpoint, or short LLM prompt Claude Code runs automatically at lifecycle points (PreToolUse, PostToolUse, Stop, SessionStart, …). Lives in settings.json, not the chat — the model can't skip it. Ten real configs with verdicts and failure receipts: Claude Code hooks.
HTML-ization: Shipping every deliverable — report, pitch, audit, deck, model — as a live interactive HTML artifact on a deployed link instead of a dead PDF/slide attachment. It gets opened, clicked into, forwarded; it stays current because it's re-rendered, not re-attached. See HTML-ization.
Inference: The act of running a model on an input to produce an output. Each chat turn is one inference.
Instance: A single context window with a system prompt, tools, and history. When the window closes, the instance dies.
Interactive mode: The default. Every Edit/Write/Bash/WebFetch shows a preview and asks "approve?" Slow on purpose — the slowness is the safety margin.
Knowledge cutoff: The date past which the model can't reliably answer. Mid-2025 for current frontier models.
/loop: /loop [interval] [prompt] — interval-driven, no evaluator. Run this prompt every N minutes (polling). The slash-command replacement for rigging cron.
LSP: A Language Server Protocol server wired into CC for symbol-level navigation ("go to definition") instead of string grep. Not automatic — install the code-intelligence plugin + a language-server binary.
MCP: Model Context Protocol. The open standard that lets any AI client talk to any tool. USB-C for AI.
.mcp.json: The repo-root file declaring MCP servers, committed so the whole team (and both Codex + CC) gets the same connector set.
MCP server: The tool side of MCP — the thing exposing Slack messages or Stripe charges. Same object a consumer UI labels a "connector," seen from the build side.
Memory curation: Keeping the memory layer clean across hundreds of sessions — deduping, verifying each lesson against its source, and holding the index under its bloat ceiling. The failure modes are a confident duplicate write and quiet abandonment.
Multimodal: Models that handle more than text — images, audio, video, code.
Mythos: Vlad's private name for Claude Opus 3 — the model his team had learned to write against, retired by Anthropic on January 5, 2026. The Mythos lesson: any model you build a moat around is on someone else's calendar. Treat as a tail risk; eval-suite the alternative before the deprecation email arrives, not after. See The Mythos lesson. Not the same thing as Claude Mythos 5, the model Anthropic shipped in June 2026 — three uses of one name, kept apart on purpose.
Mythos 5: The raw twin of Fable 5 — same underlying model, safeguards lifted in some areas, restricted to Project Glasswing partners and (next) vetted biology researchers. Reported scores sit within 1–3 points of Fable 5 except where the safeguards bite. Not Vlad's "Mythos" (that was Opus 3), and not Mythos Preview (the withheld March 2026 disclosure). See Fable 5 vs Mythos 5.
Mythos-class: Anthropic's tier above Opus — currently Fable 5 (generally available) and Mythos 5 (gated). Mythos-class traffic carries a limited 30-day retention requirement (misuse detection only, not training) and the classifier-plus-fallback safety architecture. See Fable 5.
Night shift: A monitoring agent (typically Codex) running 24/7 against Sentry, issues, and cron failures — opens a PR, posts a summary, goes back to watching. The day driver (Claude Code) reviews and merges.
Permission mode: The settings layer evaluated deny → ask → allow across managed/user/project files. Deny wins and is sticky upward. Tool names are case-sensitive.
Persona agent: An agent that writes/posts in one specific human's voice behind an approval gate. Default is always wait, never time-based auto-post, with the Four NEVERs (deals, hires, firings, condolences) off-limits.
Plan mode: The agent describes what it WOULD do — every file, every line — without writing a byte. Approve the plan as one unit. The safety stop between Interactive and Auto that nobody uses.
Plugin: A bundle of skills, MCP servers, commands, and hooks. Install once, get many capabilities.
PostToolUse: A Hook event that fires after a tool call succeeds. The format-on-save / lint / test / notify slot.
PreToolUse: A Hook event that fires before any tool call. Validate, block, or audit-log; a non-zero exit blocks the call and the stderr goes back into the model's context.
Programmatic SEO: Generating many pages from a template plus a dataset (one URL per city, tool, or query). Durable when each page carries genuinely unique data; the churn-and-burn version — boilerplate that differs only by a swapped variable — is what Google's scaled-content-abuse policy deindexes.
Prompt caching: Not a feature you turn on — a contract about what changes between calls. A cache read costs ~1/10th of full input; one CLAUDE.md edit voids the prefix and the bill spikes downstream.
Prompt injection: When adversarial text in tool output tricks an agent into doing something unintended. The new XSS.
Propose-only: A tool that can read and recommend but structurally cannot write. The safety ceiling, not just the default — the most a propose-only dreaming pass can do is surface a review file a human acts on.
Quantization: Compressing a model's weights to run on smaller hardware. Q4 quant of a 70B is shockingly close to full precision.
RAG: Retrieval-Augmented Generation. Pull relevant chunks from a vector DB, stuff into context, generate.
Sandbox: An isolated environment (Docker, VM, devcontainer) where an agent can run with skip-permissions safely. Rebuild the container if anything goes wrong.
Self-Audit: The multi-agent code audit pointed at the agent's own configuration — skills, hooks, memory, permissions, schedulers — with an adversarial red-team re-verifying every finding before anything executes. Config rot is silent by construction; the audit is what makes it visible. Method and receipts: The Self-Audit.
Skill: A folder with a SKILL.md the model auto-loads when its description matches your request. Procedural memory.
Sovereign stack: A second LLM stack you run yourself — open-weights models (GLM-4.7, Kimi K2.5, Qwen 3.5, DeepSeek) on hardware you own, behind a runtime you control (Ollama / LM Studio / vLLM). Not a replacement for Claude / GPT / Gemini — insurance against the deprecation calendar, plus a 10–30× cheaper batch tier for evaluators, summarization, and RAG. See The Sovereign Stack.
stdio: An MCP transport where the server runs as a local subprocess piping JSON-RPC over stdin/stdout. Zero network exposure — the filesystem/local-DB default.
Stop hook: A Hook that fires when the agent's turn ends. The custom-logic autonomous-loop primitive — the next turn runs if your script returns non-zero, stops on 0. Determinism beats vibes.
Subagent: A child instance spawned from your main session: own context window, own tool allow-list (restrict it to bound blast radius), own system prompt, returns one summary.
Swarm: Multiple subagents running in parallel. Fan-out / fan-in is the default pattern.
System prompt: The instruction that boots up an instance. Sets identity, constraints, defaults.
Token: About 0.75 of a word. Models count by tokens. So do bills.
Tool use: The model's ability to call functions/connectors during a turn.
ultracode: The top notch of the Opus 4.8 /effort dial — xhigh reasoning plus workflows (parallel agents, a deeper plan). A Claude Code setting, not a model effort level: a mode, not a volume knob. See Dynamic Workflows.
Vault: Your second brain. A folder of markdown files an AI can read, link, and update. Obsidian is the typical home.
Webhook: An HTTP callback. Useful for event-triggered scheduled tasks.
Worktree: Git's built-in way to check out multiple branches of the same repo into separate folders. The single most useful Unix trick when running parallel CC sessions.
Zero-click: A search that ends without a click — the answer was already on the results page or in an AI Overview. Now the default for most searches, which is why clicks are necessary-but-not-sufficient and the real KPI is conversions per source cohort.

Vocabulary you'll need.

A

B

C

D

E

F

G

H

I

K

L

M

N

P

Q

R

S

T

U

V

W

Z

The next edition lands when this list says it does.