The terms that actually matter when you run marketing in Claude. Defined plainly, with the contempt where it is earned. If a definition sounds like a vendor wrote it, it is wrong on purpose so you can see the difference.
What Claude produces when you ask it a question the data cannot support and give it no way to decline. The model defaults to maximally helpful, so it answers anyway, at the same confidence as everything else. The output reads as authoritative and is structurally untrustworthy. The fix is a refuse-condition, not a better model.
What you get when you ask Claude to "write in the voice of [famous founder]." It pulls the quotable surface tics, short sentences, contrarian framing, and staples them onto your content. The result reads as derivative because it is. Voice that compounds is your own, extracted from real samples, not a costume borrowed from someone with a bigger audience.
Putting useful content behind a form so you can market to the person later. The polite version of "your attention is the price." This site has none. Four prompts, fifteen posts, six essays, two tools, all open. The position is not generosity. It is that a fair exchange does not start with a toll.
A brand voice document made of adjectives. "Direct, confident, approachable. Professional but warm." Every brand gets roughly the same paragraph because adjectives describe voice instead of capturing it. It feels personalised and applies to everyone, exactly like a horoscope. A real voice guide is built from pattern measurements and paired contrast, not vibes.
A structural move in a prompt that measurably changes the output. There are six that matter: output spec, refuse-condition, real source material, conversational register, uncertainty flagging, and the skip condition. Everything in a good prompt library is some arrangement of these. There is no seventh secret one on a poster somewhere.
See: Claude doesn't have commands. It has levers. · Audit your own prompt
A 4,000-word system prompt that promises to research, plan, write, and analyse anything. It is five mediocre prompts mashed into one. You cannot test it, because you never know which part failed. You cannot iterate it, because changing one section moves the others. Small prompts that each do one thing, chained when needed, beat it on every axis that matters.
Telling Claude the shape of the answer before it writes one. Columns, word counts, sections, order. Without it, the model guesses at the average shape, and the average is generic. "Analyse this market" has no spec. "Five-row table, then a 150-word summary, then three recommendations ranked by effort" does. The single highest-leverage edit most prompts are missing.
The line the rest of the site is built on. The marketer is the protagonist. Claude is a tool they pick up and put down deliberately. Pricing reflects it: flat team license, no per-seat, because charging more when a team invites more people in punishes the exact behaviour good tools should reward. Every other definition in here is downstream of this one.
Charging per person who touches the tool. It taxes a team for adopting good infrastructure widely, which is the opposite of what you want to encourage. The Hub is a flat team license on purpose. The whole team uses it for one price. Inviting the junior in costs nothing extra, because it should not.
Before any work starts, you give Claude the brief and ask it to write the post-mortem of the project assuming it already failed, dated six months out, naming the specific failure. It surfaces the unsaid risk while it is still inexpensive to act on. The clients who react badly to a pre-mortem are usually the failure mode it named.
An explicit instruction telling Claude when not to answer. "Refuse to declare a trend from fewer than two data points." "If the data does not support a conclusion, say so instead of inventing one." It clamps the model's helpfulness gradient. One refuse-condition removes most of what people call hallucination, because most of it was the model being agreeable about something it could not know.
How you talk to Claude, which Claude mirrors back. Search-query input produces search-query output. A brief written the way you would brief a smart colleague produces a colleague-quality draft. The slash-command cheat sheets train the wrong register, then their users blame the model for the result.
The belief that typing /analyze does something different from typing "analyze" in the Claude chat app. It does not. There is no command parser. The slash is decorative. Real slash commands exist in exactly one place, Claude Code, the CLI tool, and that set is documented by Anthropic and changes with releases. A screenshot of 90 commands was stale the day it was posted.
The instruction that lets Claude tell you the task itself is a bad idea. "If the SERP is unwinnable, say so instead of writing the brief." Prompts written by people who actually use them include a way to say do not do this. Prompts written to look impressive never do.
The compounding cost of shipping AI output without judgment. It is paid in trust the audience withdraws, in voice sameness across a whole category, in the cleanup time your best people spend fixing what looked fine. The team that automated without taste does not see the tax, because the audience pays it first.
See: Essay: The slop tax · Run the Slop Sniffer on your own copy
Extracting a brand voice from real samples of the brand's own writing, using paired contrast and pattern measurement, so a new writer or a Claude Project can reproduce it. The opposite of adjective-based transfer, which produces a costume or a horoscope. The test is downstream: hand the guide to someone else and see if their draft sounds like the brand.
A prompt library built by saying yes to every prompt that sounds plausible. Every category gets something, every search query gets a prompt aimed at it, and half of it produces slop the buyer cannot tell apart from the half that works. A curated library is a refused library. The discipline is what makes it worth paying for.
The dictionary is the vocabulary. The library is the thing built in it: 30 prompts, each one on all six levers, none of them in the yes pile. Four are free. Start there.