The weekly note · August 1, 2026
The Words You Keep Seeing, Decoded
Seventy-nine words you will meet this week in AI newsletters, articles, and product releases, each with what it is, where it came from, and whether knowing it changes anything you do. Plus the 26 Claude Code commands worth learning first.
The vocabulary in AI writing now comes from three separate places at once: AI research, software development, and vendor vocabulary. This is a reference for 79 of the terms you will actually meet, each with a plain definition, the category it belongs to, and how it shows up in practice, plus the 26 Claude Code commands worth learning first. The filter is whether knowing a word changes something you do. The terms that fail that test are named at the end and left out on purpose.
From my desk, August 1.
I work with Claude from two sides. Teaching and coaching leaders and their teams on it, and building with it myself, most days, on client projects and on my own company's systems.
Those two feed each other. The building is where I find out what actually holds up. The teaching is where I find out which parts are hard to say out loud. And the friction that shows up most, in every room and on every project, is not the technology. It is the vocabulary.
Somebody sharp and senior is tracking the whole conversation, and then a word goes past. Context. Token. MCP. Repo. The idea underneath it is almost always simple. The word is just doing a poor job of carrying it.
So I pulled together the ones that come up most and wrote them out for my readers. Plain English, what each one is, and whether it changes anything you actually do. Not the full dictionary. The working set.
That makes this one two things at once, on purpose. It is a Signal Note, because putting the list together turned up something about this vocabulary I did not expect. And it is a reference guide, built to be bookmarked and returned to rather than read once and closed.
The unexpected part first, because it explains why the words feel harder than the ideas underneath them.
These words did not come from one place. They came from three rooms that never used to talk to each other, and now all three ship into your inbox before breakfast.
Room one is AI research. Token, inference, context window, embedding, hallucination. These are words from papers, and they carry precise meanings that survived the trip into marketing copy mostly intact.
Room two is software development, and this is the one nobody warned anybody about. Repo. Commit. Branch. Push. Shell. Terminal. These words are forty years old and they were never meant to leave engineering. They left anyway, because the tools that use them landed in everyone else's hands. A marketing director in 2026 reads the word "commit" in a newsletter and is expected to know it.
Room three is vendor vocabulary, and it moves fastest of all. Plan mode. Skills. Cowork. Effort. Connectors. One of those is a product and the rest are features and settings, which is part of why they blur: they all arrive sounding equally official. They are barely a year old, some of them barely a quarter, and half of them will be renamed before they are widely understood.
Three vocabularies, one inbox, no memo. That is the whole difficulty, and once you see it, the words stop feeling like a wall and start feeling like three shorter lists.
What made the cut, and the test I used
There are more than three hundred terms I could have put here. Most of them would be noise.
So every word had to pass one test: does knowing this change something you do? Not something you could say at a conference. Something you would actually do differently on a Tuesday, in your own business, with your own money.
Token passes, because it is how you are billed. Context passes, because it explains most of the times the AI got something wrong. Permissions passes, because it is the single most consequential setting in any of these tools and almost nobody touches it deliberately.
A few earn their place a second way. They will not change a decision by themselves, but you have to decode them to follow the sentence they are sitting in. Compute and embedding are both like that. That is a tax too, and it is worth ten seconds to stop paying it.
Transformer clears neither bar. Neither does RLHF. You will see both this month, they are real, you can read straight past them without losing the thread, and no decision you make turns on them. Those are named at the end so you know they were considered and left out on purpose, not missed.
Seventy-nine words made it. They are alphabetical, so this works as a reference rather than a read. The category tells you which room a word came from, or what job it does, which is usually the fastest way to understand why it sounds the way it does.
A through C
| Term | Category | What it is | How it shows up |
|---|---|---|---|
| TermAEO / GEO | CategoryMarketing | What it isAnswer Engine Optimization and Generative Engine Optimization. Two names for one job: earning your way into the answers AI tools give. | How it shows upTreat them as synonyms. Judge the work being proposed, not which acronym the agency picked. |
| TermAgent | CategoryAgent | What it isA model given tools, a goal, and permission to take several steps on its own instead of answering once and stopping. | How it shows upThe hardest-working word in AI writing right now. When you read it, ask what tools it can reach and what it may do without asking. |
| TermAgentic | CategoryAgent | What it isThe adjective. Software that picks its own next step rather than following a fixed script. | How it shows up"Agentic workflow" means the software chooses the order of operations. Real, and also the most oversold word in the category. |
| TermApex domain | CategoryDev word | What it isThe root of a domain with nothing in front of it. yourcompany.com, not www.yourcompany.com. | How it shows upYou will meet it in DNS, the address system behind every website, and in site launches, where the apex and the www address get configured separately. |
| TermAPI | CategoryPlumbing | What it isApplication Programming Interface. The doorway one piece of software uses to talk to another. | How it shows up"We have an API" means your other tools can send it instructions and get answers back with nobody clicking anything. |
| TermBash | CategoryDev word | What it isA command language for the terminal, and the default on most Linux systems. Modern Macs default to zsh, which is close enough that most commands you are handed run in either. | How it shows upWhen an AI tool says it is running a Bash command, it is typing into the same terminal you would. |
| TermBenchmark | CategoryModel | What it isA standardized test used to compare models against each other. | How it shows upEvery model launch leads with them. Read them as a floor, not a forecast. Benchmarks measure the test, and your work is not the test. |
| TermBranch | CategoryDev word | What it isA parallel copy of a project you can change freely without touching the version everyone else is using. | How it shows upThe safety net behind letting AI try something. Work happens on a branch, gets reviewed, then joins the main copy. |
| TermCheckpoint / rewind | CategoryClaude | What it isA saved point you can return to, rolling back the conversation, your files, or both. | How it shows up/rewind in Claude Code. The most underused safety feature there is, and it makes bold experiments cheap. |
| TermClaude Chat | CategoryClaude | What it isThe conversation surface at claude.ai. | How it shows upThinking, drafting, asking. It reaches your files only where you attach them, or where you have connected an app. |
| TermClaude Code | CategoryClaude | What it isAnthropic's agentic tool that reads and edits real files, runs commands, and works across many steps. Runs in the terminal, the desktop app, the browser, and inside code editors. | How it shows upThe one people mean when they say Claude actually did the work. Despite the name, it is not only for code. |
| TermClaude Cowork | CategoryClaude | What it isClaude working directly on your files, folders, and connected apps for work that is not code. Desktop and web, with mobile in beta. | How it shows upAnthropic's own framing is Claude Code for the rest of your work. Reorganize a folder, turn receipts into a spreadsheet, draft from scattered notes. |
| TermClaude Design | CategoryClaude | What it isA surface at claude.ai/design for producing interface designs, which can be pointed at your own component library. | How it shows upOnce synced, the designs come back using your real buttons and cards instead of generic ones. |
| TermCLAUDE.md | CategoryClaude | What it isA plain text file of standing instructions that loads at the start of every session in that project. | How it shows upWhere the rules you would otherwise repeat daily go: tone, stack, conventions, and what never to do. |
| TermCLI | CategoryDev word | What it isCommand Line Interface. Software you drive by typing rather than clicking. | How it shows upThe terminal version of a tool, usually faster and more capable than the version with buttons. Which is why serious users end up there. |
| TermCold start | CategoryDev word, Agent | What it isTwo meanings. In infrastructure, the delay when a dormant service wakes up. In AI work, beginning a session with none of the prior context. | How it shows upThe second one costs you weekly. A cold start primer is a short document that brings a fresh session fully up to speed in one read. |
| TermCommit | CategoryDev word | What it isA saved, labeled checkpoint of a set of changes, with a note explaining what changed and why. | How it shows upThe unit of work in version control. Small, frequent commits are what make AI-assisted work reviewable instead of a mystery. |
| TermCompaction | CategoryPlumbing, Claude | What it isSummarizing a long conversation so work can continue without losing the thread. | How it shows up/compact in Claude Code. What it keeps and what it drops decides whether your next hour is productive. |
| TermCompute | CategoryPlumbing | What it isRaw processing power, almost always meaning time on the graphics chips that run AI models. | How it shows upWhen an article talks about compute, it means the physical constraint under everything: how much silicon is running, and who got there first. |
| TermConnector | CategoryPlumbing | What it isA prebuilt integration linking an AI tool to an outside app such as Gmail, Drive, or a CRM. | How it shows upThe friendly wrapper. Most connectors are MCP servers underneath with the hard setup already done. |
| TermContext | CategoryModel | What it isEverything the model can see right now: your message, the conversation so far, files, instructions, and tool results. | How it shows upThe most important word on this list. Almost every time the AI gets something wrong, it is a context problem rather than an intelligence problem. |
| TermContext engineering | CategoryModel, Plumbing | What it isThe deliberate work of deciding what goes into the context and what stays out. | How it shows upThe skill that quietly replaced prompt engineering. What you leave out matters as much as what you put in. |
| TermContext window | CategoryModel | What it isThe size limit of the context, measured in tokens. | How it shows upA one-million-token window is roughly a shelf of books. Bigger is not automatically better, because a full window of noise still produces noise. |
| TermCrawl / crawler | CategoryMarketing, Plumbing | What it isAutomated software that reads pages and follows links. | How it shows upSearch engines crawl to index. AI companies crawl to train and to answer. Whether you allow either is a decision, and most sites make it by accident. |
D through M
| Term | Category | What it is | How it shows up |
|---|---|---|---|
| TermDeploy | CategoryDev word | What it isPushing a change out to where real people can see it. | How it shows up"It works on my machine" and "it is deployed" are different claims. Verify the live address, not the preview link. |
| TermDiff | CategoryDev word | What it isA line-by-line comparison of before and after. | How it shows upHow you review AI-written work without reading everything. Read the diff, not the file. |
| TermEffort | CategoryClaude, Model | What it isHow hard the model works on a single turn. Low, medium, high, xhigh, max, and ultracode. | How it shows upA direct trade of speed for depth. Most people leave it at default and never find out what the model can do. |
| TermEmbedding | CategoryPlumbing | What it isText turned into a long list of numbers so software can compare meaning instead of matching words. | How it shows upThe engine under semantic search. Two sentences with no words in common can end up sitting right next to each other. |
| TermEval | CategoryModel | What it isA repeatable test of output quality on your own work, not a public benchmark. | How it shows upThe thing almost nobody builds and everybody needs. Without evals you are grading AI output on how it feels. |
| TermFew-shot / zero-shot | CategoryModel | What it isWhether you gave the model examples. A few examples is few-shot, none is zero-shot. | How it shows upExamples are still the cheapest quality upgrade available. Newer models need fewer than they used to, which is not the same as none. |
| TermFine-tuning | CategoryModel | What it isFurther training a model on your own data so it internalizes a pattern. | How it shows upExpensive, slow, and usually the wrong first answer. Most problems people bring to fine-tuning turn out to be context problems. |
| TermFoundation model | CategoryModel | What it isA large general-purpose model trained on broad data, built to be adapted rather than used raw. Frontier model means the current most capable tier. | How it shows upThe layer everything else sits on. Building on a foundation model means renting the hard part instead of building it. |
| TermFrontend | CategoryDev word | What it isThe part of software a person sees and touches. Backend is the part they do not. | How it shows upWhen a marketer says "the site," they almost always mean the frontend. When a developer says it, they may not. |
| TermGit | CategoryDev word | What it isThe version control system nearly everyone uses. Tracks every change, who made it, and lets you go back. | How it shows upNot the same thing as GitHub, which is a company that hosts Git projects. Git is the tool, GitHub is the address. |
| TermGraph / knowledge graph | CategoryPlumbing | What it isInformation stored as things and the links between them, rather than as rows in a table. | How it shows upHow memory systems stay useful as they grow. A fact links to related facts, so retrieving one surfaces the neighborhood around it. |
| TermGrounding | CategoryAgent, Model | What it isTying an answer to a real, checkable source instead of the model's recollection. | How it shows upThe practical antidote to hallucination. "Grounded in your documents" means it read them, rather than remembered something like them. |
| TermGuardrails | CategorySafety, Agent | What it isRules limiting what a system may do, enforced outside the model's own judgment. | How it shows upReal guardrails are permissions and code. A line in a prompt asking politely is not a guardrail. |
| TermHallucination | CategoryModel | What it isConfidently stated output that is simply wrong. | How it shows upThe name is generous, because the behavior is fluent invention. Verify anything you are about to forward. |
| TermHomebrew (brew) | CategoryDev word | What it isThe package manager most Mac users install command-line software with. One command fetches a tool and everything it needs to run. | How it shows upbrew install is how most of the tools in this note get onto a Mac in the first place. If a setup guide starts with a brew line, this is why. |
| TermHook | CategoryClaude, Plumbing | What it isCode that fires automatically at a set moment, such as before a file is edited or when a session starts. | How it shows upHow you enforce a rule instead of hoping for one. Formatting, tests, and notifications all live here. |
| TermHuman in the loop | CategoryAgent | What it isA required human approval at a specific point in an otherwise automated process. | How it shows upThe phrase is carrying a lot of weight in AI governance documents. When you see it, ask exactly which step and exactly which human. |
| TermInference | CategoryModel | What it isRunning a model to produce an answer, as opposed to training it. | How it shows upYour monthly bill is almost entirely inference. Training is the factory. Inference is every unit coming off the line. |
| TermKnowledge cutoff | CategoryModel | What it isThe date after which the model was not trained on anything. | How it shows upWhy a model can be brilliant and still not know what happened last month. Give it a search tool, or give it the document. |
| TermLLM | CategoryModel | What it isLarge Language Model. The kind of model under nearly every AI tool you use. | How it shows upIncreasingly a lazy shorthand. When it matters, name the actual model and version, because capability varies enormously inside that one acronym. |
| TermLoop | CategoryAgent | What it isThe cycle an agent runs: plan, act, check the result, adjust, repeat until done. | How it shows upWhether there is a loop is the difference between a chatbot and an agent. It also names Claude Code's /loop, which reruns a prompt on an interval. |
| TermMCP | CategoryPlumbing | What it isModel Context Protocol. An open standard for connecting AI tools to outside systems, so an integration written once works across tools. | How it shows upThe reason your AI tool can read a ticket, query a database, or post to Slack. An MCP server is one such connection. |
| TermMemory | CategoryModel, Claude | What it isFacts kept between sessions so you are not reintroducing yourself every time. | How it shows upDifferent from context. Context is this conversation. Memory is what survives it. |
| TermModel | CategoryModel | What it isThe trained system itself. Opus, Sonnet, and Haiku are models. Claude is the family they belong to. | How it shows upNaming the model matters more than people think. "I asked Claude" tells you far less than "I asked Opus 5 at high effort." |
| TermMultimodal | CategoryModel | What it isAble to work with more than text: images, audio, documents, video. | How it shows upNow table stakes. The practical version is that you can paste a screenshot and ask what is wrong with it. |
N through Z
| Term | Category | What it is | How it shows up |
|---|---|---|---|
| TermOpen weights | CategoryModel | What it isThe trained values inside a model are published, so anyone can download and run it themselves. | How it shows upFrequently called open source, which is usually inaccurate. Open weights rarely means the training data or the code came with it. |
| TermOrchestration | CategoryAgent | What it isCoordinating several models, agents, or steps into one reliable sequence. | How it shows upWhere most of the real value gets created, and where most projects fail. The hard part was never the model. |
| TermParameters | CategoryModel | What it isThe learned numerical values inside a model. The 7B or 70B in a model's name counts them in billions. | How it shows upA rough proxy for size, and a poor one for quality. Training data and method beat raw count regularly now. |
| TermPermissions | CategoryClaude, Safety | What it isThe rules governing which actions a tool may take on its own and which need your approval. | How it shows upThe most consequential setting in any agentic tool. Decide it deliberately, once, early. /permissions in Claude Code. |
| TermPlan mode | CategoryClaude | What it isA mode where Claude researches and proposes an approach, changing nothing until you approve it. | How it shows up/plan, or Shift+Tab twice. One press lands you in accept-edits instead, which is the opposite of what you want here. This is the habit that separates people who trust the output from people who audit it afterward. |
| TermPlugin | CategoryClaude | What it isA shareable bundle of skills, agents, hooks, and connections a team can install together. | How it shows upHow one person's workflow becomes the team standard. Installed from a marketplace, versioned like software. |
| TermPR (pull request) | CategoryDev word | What it isA proposed set of changes, opened for review before it joins the main version. | How it shows upThe review checkpoint. In AI-assisted work it is where a human reads the diff and decides. |
| TermPrimer | CategoryAgent | What it isA short document written to bring a fresh session, or a new person, fully up to speed in one read. | How it shows upPairs with cold start. A good primer is the difference between resuming work and restarting it. |
| TermPrompt | CategoryModel | What it isWhat you send the model. | How it shows upPrompt engineering as a discipline has quietly deflated as models improved. Being specific about the goal still beats being clever about the phrasing. |
| TermPrompt injection | CategorySafety | What it isHostile instructions hidden inside content the model reads, meant to hijack what it does next. | How it shows upThe real security concern in agentic tools. If your agent reads the open web or your inbox, assume something out there is trying it. |
| TermPush | CategoryDev word | What it isSending your saved commits from your machine up to the shared copy. | How it shows upCommitting saves locally. Pushing publishes. Work that is committed but never pushed exists on exactly one computer. |
| TermRAG | CategoryPlumbing | What it isRetrieval Augmented Generation. Fetching the relevant documents first, then answering from them. | How it shows upThe standard pattern for "answer from our own content." Less fashionable than it was two years ago, and still the right answer often. |
| TermRate limit | CategoryPlumbing | What it isA cap on how many requests, or how much text, you can send in a given window. | How it shows upThe wall you hit the moment automation scales. Worth knowing before you design a process around a tool, rather than after. |
| TermRepo | CategoryDev word | What it isShort for repository. The folder holding a project plus its entire history of changes. | How it shows up"It's in the repo" means it is in the tracked project, not sitting on somebody's desktop. |
| TermRubric | CategoryModel | What it isThe specific qualities you are grading for, named before you look at the output. | How it shows upWhat an eval is actually made of, and the difference between a score you can defend and "that reads fine." It is also how you get useful grading out of AI: hand it the rubric and you get a judgment instead of a compliment. |
| TermSandbox | CategorySafety, Plumbing | What it isAn isolated environment where code can run without touching anything real. | How it shows upHow you let an agent execute something nobody has read yet. If a vendor cannot describe theirs, that is your answer. |
| TermShell | CategoryDev word | What it isThe program that interprets what you type in the terminal. Bash and zsh are shells. | How it shows upTerminal is the window, shell is what listens. People use the words interchangeably and it almost never causes harm. |
| TermSkill | CategoryClaude, Agent | What it isA packaged set of instructions for one kind of task, loaded only when it is relevant. | How it shows upFollows the open Agent Skills standard. The practical unit of teaching AI how your company does something, without paying for it in every conversation. |
| TermSlop | CategoryMarketing | What it isGeneric, high-volume AI output published without judgment. | How it shows upNow a load-bearing word in editorial and marketing conversations. It describes an effort problem rather than a tool problem. |
| TermSubagent | CategoryAgent | What it isA separate agent handed a scoped task, which reports back a result instead of its whole process. | How it shows upHow one session handles work that would never fit in a single context. The main thread keeps the conclusion, not the raw material. |
| TermSuperpowers | CategoryClaude, Agent | What it isA widely used open-source plugin by Jesse Vincent that adds a library of development skills to Claude Code. | How it shows upCommunity-built, not an Anthropic feature, which is the thing to check before you install any plugin. It runs with whatever permissions you have given the tool. |
| TermSystem prompt | CategoryModel | What it isThe standing instructions a product sends the model before your message ever arrives. | How it shows upEvery AI product has one, and it shapes everything. When two tools running the same model behave differently, this is usually why. |
| TermTerminal | CategoryDev word | What it isThe text window where you type commands straight to your computer. | How it shows upIt is where the most capable version of most AI tools lives, which is why so many people who never wanted a terminal now have one open. |
| TermToken | CategoryModel | What it isThe unit models read and bill in. Roughly three quarters of a word in English. | How it shows upEverything is priced, limited, and measured in tokens. A page of text runs about 650 of them. |
| TermTool use | CategoryAgent | What it isA model calling actual software mid-answer: a search, a database, a calculator, your CRM. | How it shows upAlso called function calling. It is what turns a model from something that talks into something that does. |
| Termultracode | CategoryClaude | What it isThe top effort setting, where Claude plans, fans work out, and verifies far more before answering. | How it shows upSession-only and expensive by design. Reach for it when being wrong is costly and being slow is not. |
| TermVector database | CategoryPlumbing | What it isA database built to store embeddings and find the nearest matches fast. | How it shows upThe storage layer under most RAG systems. Pinecone, Weaviate, and pgvector are the names you will see. |
| TermVibe coding | CategoryMarketing, Dev word | What it isBuilding software by describing what you want and accepting what comes back without reading it closely. | How it shows upCoined in early 2025 and now used both admiringly and as an insult. Fine for a prototype, a liability in production. |
| TermWorkflow | CategoryAgent, Claude | What it isA defined sequence of steps, increasingly one that runs several agents in a deliberate order. | How it shows upIn Claude Code, /workflows shows a multi-agent run live. Worth saying plainly: a pile of tools is not a workflow. |
| TermWorktree | CategoryDev word | What it isA second working copy of the same project, checked out to a different branch, in its own folder. | How it shows upHow several agents work in parallel without overwriting each other. Isolation without cloning the whole project. |
The five Claude surfaces, since the names blur
One recommendation, since the terminal keeps coming up
This one is not a term, so it does not belong in the table. It comes up in almost every session anyway, because once somebody accepts that the terminal is where the good version of these tools lives, the next question is always which terminal.
On a Mac I use Ghostty. It is free, open source, and made by Mitchell Hashimoto, who co-founded HashiCorp. Three things make it worth the ten minutes it takes to switch.
Fair framing, since Hashimoto is careful about this himself: he does not claim it is the fastest or the most capable terminal available. He set out to build one that is competitive on speed, features, and native feel at the same time, rather than strong at two and weak at the third. That is exactly why I recommend it to people who are not looking for a hobby. It is one install and then you stop thinking about it.
brew install --cask ghostty if you have Homebrew. The download on their site works fine if you do not. It needs macOS 13 or later, and if your machine is managed and installs are locked down, the Terminal that came with your Mac runs everything in this note.
The commands worth learning first
Claude Code ships with 103 commands as I write this, and the number keeps climbing. You do not need 103. These 26 are the ones that change how a week goes, grouped by what they are for rather than alphabetically, because that is how you will reach for them.
Managing what the model can see
| Command | What it does | When to reach for it |
|---|---|---|
Command/context | What it doesShows what is currently filling the context, as a colored grid. | When to reach for itWhen answers start drifting. It almost always reveals something bloated you forgot was loaded. |
Command/compact | What it doesSummarizes the conversation so far to free room. Takes instructions on what to preserve. | When to reach for itBefore a long session runs out of room. Tell it what to keep. |
Command/clear | What it doesStarts fresh with empty context, leaving the old conversation resumable. | When to reach for itBetween unrelated tasks. Carrying yesterday's context into a new job is the most common self-inflicted wound there is. |
Command/resume | What it doesReopens an earlier conversation. | When to reach for itMonday morning. It picks up exactly where you stopped. |
Command/rewind | What it doesRolls the conversation and your files back to an earlier point. | When to reach for itThe moment something goes sideways. Also answers to /undo. |
Controlling how hard it works
| Command | What it does | When to reach for it |
|---|---|---|
Command/model | What it doesSwitches models and sets your default. | When to reach for itWhen the work changes shape. Press s in the picker to switch for this session only. |
Command/effort | What it doesSets how hard the model works per turn, from low through max, plus ultracode. | When to reach for itTurn it up for architecture and review. Turn it down for mechanical edits. |
Command/fast | What it doesToggles fast mode, which keeps you on Opus and speeds up output. As I write this it is a research preview and costs more per token. | When to reach for itIterative back-and-forth where waiting breaks your rhythm. Check the current pricing before you make it a habit: on a subscription it has been drawing on usage credits rather than your plan's included limits. |
Command/plan | What it doesEnters plan mode. Claude proposes an approach and changes nothing until you approve. | When to reach for itAny task where you want to see the thinking before the edits. Shift+Tab twice does the same thing. One press does not. |
Command/goal | What it doesSets a condition Claude keeps working toward across turns until it is met. | When to reach for itLong jobs with a clear finish line. |
Checking the work
| Command | What it does | When to reach for it |
|---|---|---|
Command/code-review | What it doesReviews your current changes for real bugs. Has effort levels, and --fix applies findings. | When to reach for itBefore you ship anything. Passing ultra runs a deep cloud review of the whole branch. |
Command/security-review | What it doesReviews pending changes specifically for security problems. | When to reach for itBefore anything touching login, payments, or customer data goes out. |
Command/simplify | What it doesRuns four agents in parallel looking for cleanup: reuse, simplification, efficiency, and whether the code sits at the right level. | When to reach for itAfter it works but before you call it done. Quality only, it does not hunt bugs. |
Command/verify | What it doesBuilds and runs your actual app to confirm a change works, rather than trusting tests. | When to reach for itWhenever "the tests pass" is not the same as "I watched it work." |
Command/doctor | What it doesA full setup checkup that finds problems and offers to fix them. | When to reach for itQuarterly, and any time something feels off. It also trims bloated instruction files. |
Teaching it your rules
| Command | What it does | When to reach for it |
|---|---|---|
Command/memory | What it doesEdits your CLAUDE.md files and manages what is remembered between sessions. | When to reach for itThe moment you catch yourself repeating an instruction. Write it once instead. |
Command/init | What it doesWrites a starting CLAUDE.md for a project by reading the project itself. | When to reach for itDay one in any new project. |
Command/config | What it doesOpens settings, or sets one directly with key=value. | When to reach for itTheme, model, output style. Run /config --help to see every key. |
Command/permissions | What it doesManages which tools may run without asking you first. | When to reach for itOnce, deliberately, early. It is the setting with the most consequence attached to it. |
Command/skills | What it doesLists every available skill. Press t to sort by what each one costs you in tokens. | When to reach for itWhen you want to know what is actually loaded and what you are paying for it. |
Reaching outside
| Command | What it does | When to reach for it |
|---|---|---|
Command/mcp | What it doesManages connections to outside tools and their authentication. | When to reach for itWhenever a connected tool stops responding. reconnect fixes most of it. |
Command/plugin | What it doesInstalls and manages plugins from marketplaces. | When to reach for itWhen you want a teammate's whole setup rather than one borrowed instruction. |
Command/subtask | What it doesHands a task to a background agent that inherits the full conversation and reports back. | When to reach for itResearch or a side investigation you do not want filling your main thread. |
Command/deep-research | What it doesFans out web searches, cross-checks the sources, and returns a cited report. | When to reach for itReal research questions. It is a workflow, not a single search. |
Command/btw | What it doesAsks a quick side question without adding it to the conversation. | When to reach for itAny time you need a fact but do not want to pollute the context you have carefully built. |
Command/usage | What it doesShows session cost, plan limits, and a breakdown by skill, subagent, plugin, and MCP server. | When to reach for itWeekly. It tells you where your budget is actually going. /cost and /stats are the same command. |
What I left out, on purpose
You will meet these this month. They are real, they are correctly used, and knowing them will not change anything you do. If you want them, they are a search away. They are not in the table because the table is meant to be useful rather than complete.
Four moves, in order
- Learn three words before you learn seventy-six. Context, token, and agent carry most of the weight. Everything else is easier once those three are solid.
- Run the two commands that show you your own habits.
/contexttells you what the model is actually looking at./usagetells you what it costs. Most people have never run either, and both take ten seconds. - Write down the rules you keep repeating. Every instruction you have given twice belongs in a CLAUDE.md file or a skill. You are otherwise paying to say the same thing forever.
- Turn plan mode on before you turn autonomy up. Seeing the approach before the edits is what makes it reasonable to give a tool more room later. Do it in that order and the whole thing stays yours.
The part nobody puts in a glossary
The words are the easy part. What actually separates a team getting real work out of these tools is not vocabulary. It is whether anyone has decided how the team will work. Which tasks go to AI, who reviews the output, what the standing rules are, and where the line sits between assisted and unattended. The reviewing half of that I have already written down, rubric and all, in the loop I was already running.
The rest of it is the work I do with teams in Claude Team Training: four weeks inside one team's real week, until the work they produce every week runs with Claude and they demo it themselves.
Keep this one open in a tab. That is what it is for. The words you keep seeing are going to keep coming, and the list is going to keep growing, but the three rooms they come from are not going to change, and neither is the test. If knowing it does not change what you do, you can let it go by.
All signal. No noise.
---
Frequently asked questions
What are the most important AI terms to know in 2026?
Three carry most of the weight: context, token, and agent. Context is everything the model can see right now, and it explains almost every time the output misses. A token is the unit models read and bill in, roughly three quarters of a word, so it is how everything gets priced and limited. An agent is a model with tools, a goal, and permission to take several steps on its own. Understand those three and the rest of the vocabulary gets noticeably easier.
What is MCP in AI?
MCP stands for Model Context Protocol. It is an open standard for connecting AI tools to outside systems, so an integration built once works across different tools instead of being rebuilt for each one. In practice it is the reason an AI assistant can read a support ticket, query your database, or post to Slack. An MCP server is one of those connections. A connector is usually an MCP server with the setup already done for you.
What is the difference between context and memory?
Context is what the model can see in the current conversation: your message, the history so far, any files or instructions loaded, and the results of anything it just ran. Memory is what survives after that conversation ends. Context resets. Memory persists. Most frustration people blame on the model forgetting is a context problem, and it is fixable by deciding more deliberately what goes in.
Why do AI articles use so many software developer words?
Because the tools came from engineering and moved outward faster than the vocabulary got translated. Repo, commit, branch, push, shell, and terminal are decades-old developer words that were never meant to leave engineering. They left anyway, because the AI tools that use them landed in everyone's hands. That is why the vocabulary feels harder than the concepts: you are reading three separate vocabularies at once, from AI research, from software development, and from vendor vocabulary.
How many slash commands does Claude Code have?
There are 103 commands as of this writing, and the list grows regularly. You do not need most of them. The ones that change how a week goes are /context and /compact for managing what the model sees, /model and /effort for controlling how hard it works, /plan for seeing the approach before any edits happen, /code-review and /verify for checking the work, and /usage for knowing what it costs.
What is the difference between Claude Code in the terminal and Claude Code in the desktop app?
It is the same Claude Code underneath. What changes is how much of the work you can see, and how many things you can run at once. The terminal is one conversation in a text window, and it goes anywhere a terminal goes, including a server you connect to remotely or a scheduled job that runs with nobody watching. The desktop app puts that same engine in a real application: a sidebar of sessions running in parallel, a visual review of every change before you accept it, a built-in file editor and terminal, and a notification when a task finishes while you are off in another window. On a git project, each desktop session gets its own isolated copy of the code, so two sessions working at the same time cannot overwrite each other. The app also holds Chat and Cowork in its other two tabs, so all three surfaces live in one place. One thing worth knowing before you split your work across both: the desktop app can see and message your other desktop sessions, but it cannot see sessions you started in the terminal. If you are choosing, the terminal is the place to learn, and the desktop app is where you go the first time you want two things running at once. /desktop moves a session you already started over to the app.
What is the difference between Claude Code and Claude Cowork?
Both are Claude working on real files across many steps rather than just answering a question. Claude Code came first and is built for software work, running in the terminal, a desktop app, the browser, and inside code editors. Cowork applies the same approach to work that is not code: reorganizing files, building spreadsheets from receipts, drafting from scattered notes, and it runs on desktop, web, and mobile. Anthropic's own framing for Cowork is Claude Code for the rest of your work.
Next note · Wednesday, August 5
More from the Signal.
Want this thinking applied to your business?
Signal Notes can sharpen the thinking. A strategy call turns it into a plan.