The AI Second-Order Thinking skill for Claude

The problem the skill tackles is a structural flaw almost every plan has: the same person comes up with it, judges it and approves it. The file's frontmatter names the complaints literally: everyone is enthusiastic and that makes me nervous, I came up with this plan myself and approved it myself, we have had no pushback on this, I want to know what I am not seeing. A plan without pushback has not been tested, only repeated. This skill organises that pushback, systematically, without needing a colleague brave enough to play the part.

It is written for business owners and teams with a decision that really matters: a launch, a pitch, an investment, a price change, a new market. It is part of the skill library NL that we make available for free from the AI and automation service, with no account and no sales email afterwards.

One thing up front: the attack is an exploration, not a prediction. It does not say what will happen. It says what is plausible once other parties react to your move, and that is something else. And it ends every analysis with a section you rarely get from human advisers: where my attack was weak, with the information needed to test that part too.

The first mistake is the empty chair. The plan states what you are going to do, but not what the others do once you do it. The competitor defending their market share. The existing customer who just paid the old price and wants the difference back. The supplier who does not want their brand positioned as cheap in the market. That is why at least three opponents come to the table, each with motive, means, expected move and the timeframe in which that move happens.

The second mistake is the assumption that passes as fact. Customers choose on price. The competitor reacts slowly. Six months of cash is enough. Every assumption therefore gets a label: strong, weak or untested. The last two always come with the question of what happens if it is wrong, plus the cheapest way to find that out before the market does it for you.

The third mistake is stopping at the first effect. That is exactly where the difference lies between a list of risks and an analysis. In diff form, with the logic from the skill's example:

Next comes the method: seven things the skill always does. Ask for the plan as a decision with a date and reformulate it into one sentence. Map at least three opponents. Label every assumption. Write out the consequence chain for at least three major actions through to at least the second-order effect. Run a pre-mortem from a specific moment in the future. Deliver choices for this week, not general recommendations. And finally, name where the attack was weak, meaning which parts of the plan could not be properly tested with the information supplied.

The file contains a mandatory seven-point input checklist, from the decision with budget to the question of what has already been tried and why it did not work. Whatever is missing, it asks for all at once, and then carries on with what is there. The fixed output counts eight blocks: the plan in one sentence, the opponents, the assumptions with evidence, the consequence chains, the pre-mortem, the choices for this week, the signals to monitor, and the section where my attack was weak.

What stands out are the writing rules the skill imposes on itself: clear language addressed directly to you, no hollow superlatives, concrete over abstract, figure over adjective, and the opponent speaking in the first person where that is sharper, because a competitor stating what they are going to do is clearer than an abstract risk. There is also a separate review mode for existing documents, including an [A] marker for every sentence that is an assumption but is presented as a fact. Want to learn to build this kind of instruction file yourself? The full explanation is in writing a SKILL.md NL.

The second building block is red teaming. The structured form of pushback comes from the programme of the University of Foreign Military and Cultural Studies within the US Army, which is where the Red Team Handbook comes from: techniques for systematically attacking your own assumptions. Business strategist Bryce Hoffman translated that approach into business in Red Teaming from 2017. The third building block is the pre-mortem by psychologist Gary Klein, described in Harvard Business Review in 2007: place yourself at a moment in the future where the plan has already failed and write down why, because that produces sharper answers than the question of what could go wrong.

The file also has three variants that the skill draws on when the context calls for it. The inversion of mathematician Carl Jacobi, made popular by Charlie Munger: do not ask how you succeed but how you guarantee failure, and turn the answers into measures. Devil's advocacy, where one role is tasked with attacking everything, useful in groups. And the analysis of competing hypotheses by Richards Heuer from the intelligence world, for situations where the question is not what could go wrong but what is already going on right now. Anyone who wants that first technique as a standalone skill can find it in the Inversion Thinker skill; the pre-mortem has one too, the Pre-Mortem Analyst skill NL.

Even without installing the skill you can test your plan with these principles: write down three parties who benefit from your failure, label your assumptions honestly, and think through every action at least one order further than you had planned.

After that, using it is simple: describe your plan with a date and a budget and ask for the attack. The mandatory input checklist makes sure the skill only asks for what is genuinely missing, all at once. If you get stuck anywhere while installing, every step is described in more detail in installing Claude skills NL, and the broader explanation of working with AI is in the knowledge base.

There is a separate mode for existing work. If you hand it a launch document or an investment proposal to review, the skill marks every sentence that is an assumption but is presented as a fact with an [A], checks which parties are entirely missing from the document (usually the company's own customers, its own team, or a platform the business depends on), counts how far the document thinks things through, looks for the missing fallback options, and closes with a pre-mortem and the cheapest tests, ranked on the ratio between cost and information gained.

When do you leave it be? When there is no plan yet: the skill attacks, it does not write. And when your question is actually a different one. If you only want the answer to why a finished plan failed a year from now, the Pre-Mortem Analyst NL is enough. If you are stuck in optimism and want to flip the question, take the Inversion Thinker. And if this is not about this one move but about whether your business can withstand shocks, the Antifragile Thinker is the logical next step: first keep the plan standing, then strengthen the structure underneath it.

The sharpest second-order question for most businesses, incidentally, is not about a new plan, but about the status quo: what happens if your customers can no longer find you tomorrow, and who benefits from that? The free SEO scan shows in a few seconds where your site stands, before a competitor answers that question for you. In the guide eight decision models compared NL you can see how this model relates to the others. And if you want to talk further about what AI can do for your business, from separate skills to full automation, we are happy to simply talk about that.