What is actually in the SKILL.md
A skill is only as good as its instructions, so we simply describe them here. The file opens with concise frontmatter. The body preserves the full scope and all original trigger phrases that tell Claude when to pick up the skill. Not only for the literal requests like red team, attack my plan or be my devil's advocate, but also for sighs like: I am launching next month and I do not know what I am missing, everyone around me is enthusiastic and that makes me nervous, I made and assessed this plan myself, I want someone to tear this apart first.
Anyone who says something like that to Claude while the skill is loaded gets it automatically. The frontmatter and full scope section also explicitly rule out: inventing plans that do not exist yet, attacking people, and technical security testing.
Next comes the role switch, and it is stated bluntly: once the skill is active, Claude stops defending the plan and attacks it with the means a real adversary would have. It also says this out loud to you, so you know exactly what stage the conversation is at. It then carries out seven fixed actions: naming the role switch, mapping at least three adversaries, exposing the assumptions under the plan including the unspoken ones, scoring eight to twelve assumptions on likelihood times impact, describing concrete attack vectors per adversary, writing a pre-mortem of a few paragraphs in which the plan has failed, and delivering one defensive action per main danger that you can carry out this week.
Before anything gets attacked, the skill runs through a mandatory input checklist: the plan in the form it exists in now, what it has to deliver and by when, what it costs in money, time and attention and what is irreversible, who is affected if this succeeds, what is already fixed and what can still change, the hard deadline, and what you personally find most nerve-racking. That last question is the sharpest of the set, because what you find most nerve-racking is usually the point you do not dare look at. If anything is missing, it asks in a single round and then carries on with marked assumptions.
The output always follows the same six phases: goal and scope of the attack, the adversary map, the assumption list, the attack vectors, the damage score and finally the pre-mortem with the defence. And there is a separate review mode: feed it an existing plan, business case or risk paragraph, and you first get an assumption density with quotes, a list of what is missing from it, an optimism check on timeline, conversion and cost, and a marker per section of whether a mistake there can be undone. Irreversible steps belong late in the plan, not early. The full analysis then still follows.
Anyone who wants to train probing questions as a standalone skill can turn to the Socratic Method skill for that; anyone who wants to learn to build a file like this themselves finds the full explanation with a template at writing your own SKILL.md NL.
The file closes with writing rules that guard the tone, and one of them sums up the whole skill: Say the uncomfortable thing out loud. A red team that softens its analysis is useless. But attack the plan, never the person who made it.