What is really in the SKILL.md
A skill is only as good as its instructions, so we simply describe them here. The file opens with a frontmatter: the name ai-aida-copywriter, version 1.0.0, MIT licence, and a long description that decides when Claude picks up the skill. That trigger list is wider than you would think. Besides obvious questions such as AIDA text, sales email, cold email, Google Ads copy, writing a headline and VSL script, it also reacts to complaints: my copy does not convert, readers drop off, no clicks on my advert, nobody replies to my cold email. Review requests trigger it too, including take the AI out and make this more human.
Then comes the theory. AIDA is described as a hierarchy of effects model from 1898, first written down by Elias St. Elmo Lewis, later extended and confirmed by, among others, Strong in 1925, and used in practice by agencies such as Ogilvy and BBDO. Every letter gets its own definition: Attention is stopping the scroll among all the other stimuli, Interest is the reader recognising themselves, Desire is wanting a result that has been made concrete, and Action is the specific act at the right moment with the right ease.
The core is a list of six things the skill always does. It asks for the minimum necessary context or infers it from your input. It splits the output explicitly into the four stages. It delivers a concrete line of text per stage, so text and not advice. It delivers an alternative variant per stage so you can A/B test. It marks per stage which principle was used, for example contrast, specificity, social proof, urgency or movement. And it closes with a list of points to watch that you have to test yourself.
Before writing anything it runs through a mandatory input checklist with six fields: your target audience as concretely as possible including their level of awareness, your offer or product, the channel, the desired action, the biggest pain point or the biggest dream, and the piece of proof you have. The instruction next to it is sharp: only ask about what is missing, then stop asking. So you do not get a form.
A striking number of rules are about language. It states that the skill writes in clear British English, addressing the reader directly and informally as you, that no dashes appear in running text, that AI is always written in capitals and that small language quirks are allowed so it sounds human. Lists of exactly three items are forbidden, because that rhythm gives the machine away. There is a banned list of AI words and turns of phrase, including in the world of, discover, unlock, seamless, effortless, dive into and in a nutshell. And there is a second list of hollow superlatives: the best, amazing, incredible, fantastic, groundbreaking. As a rule of thumb: concrete beats abstract, a number beats an adjective, and short sentences alternate with long ones.
The output structure is fixed in five blocks, always in this order. First a context summary of two to four sentences, which serves as a check on your input. Then the main variant in four blocks, with the text and a line starting with Principle for each block. Then the B variant with the same structure and a different angle. Then the points to watch per stage, one or two per stage and explicitly no fixed number. And finally a paragraph on source and justification stating that the basis comes from Lewis in 1898, and which variant was used and why.
The file closes with the review mode, the list of things the skill never does, and a source list of five titles: Lewis 1898, Strong 1925, Ogilvy 1983, Sugarman 2007 and Caples 1932. Anyone who wants to build a file like this themselves finds the structure step by step in writing a SKILL.md.