What's actually in the SKILL.md
A skill is only as good as its instructions, so we simply describe them here. At the top is a frontmatter that tells you when Claude should pick up the skill. Not just for obvious terms like mom test, Rob Fitzpatrick, customer interview and validation conversation, but also for exasperated remarks: everyone says it's a good idea, how do I know if they'd actually buy it, why isn't anyone buying when everyone's enthusiastic. Say something like that to Claude while the skill is loaded, and it kicks in automatically. The frontmatter also states when it should not trigger: for quantitative survey research with statistical requirements, and for conversations without a learning question, such as a demo or a negotiation.
A theory chapter follows, and that's unusual for a skill of this size. It explains where the name comes from, sets out the three rules, splits bad data into compliments, fluff and ideas from the other person, and closes with commitment and advancement: a conversation only counts as a result once the other person hands over time, reputation or money. That theory isn't there for show. Because the rules are in the file, Claude can name which rule each question breaks instead of just passing judgement.
The core is a fixed process of six steps. Step 1 is reading and recognising: the skill reads your list literally and feeds back in three lines what it infers from it, goal, target audience and learning question, so you can correct it straight away. Step 2 is asking follow-up questions when context is missing, with an explicit maximum of two questions. If the learning question is missing, it can't judge anything, because a question isn't good or bad in itself, only in relation to what you want to learn.
Step 3 is the diagnosis per question, with the three labels from FIG.02. Step 4 is rewriting every amber and red question, with four requirements for a good rewrite: it asks about the last time it happened and names a moment when doing so, it asks what the other person did and not what they thought, it doesn't name your solution, and it leaves room for the answer you'd rather not hear. The original question stays alongside it, deliberately, so you can see the difference and learn it yourself.
Step 5 is the three follow-up options per topic: facts, cost and commitment, set out in FIG.03. Step 6 is the compliment check, and it's compulsory. It scans your list for questions that invite compliments and describes how to respond if a compliment still crops up during the conversation: don't say thanks and move on, redirect to behaviour instead. There's a literal example sentence for this, along with the note that this step must never be skipped, not even if the rest of the questionnaire is green from top to bottom.
Besides the six steps, the file contains four further fixed parts. An input checklist of five points: the questionnaire, the target audience as specific as possible (with the note that Entrepreneurs is not a target audience), the learning question, the stage you're at, and whether you also want to sell something in this conversation, because then different rules apply. An output format with fixed headings, including a diagnosis table, a was-and-becomes table, the follow-up options per topic, the compliment check and a closing block that explicitly states what you know after this conversation and what you don't. A shortened example, the photographer example from FIG.01. And finally two chapters that guard the boundaries: eight things the skill never does, and six honest limits of the framework itself.
If you want to write a file like this yourself, or adapt this one to your own trade, that's the whole point: it's released under an MIT licence. Writing a SKILL.md NL explains how to build a file like this, from frontmatter to refusals block.