The AI Inversion Thinker skill for Claude

Inversion explores plausible failure mechanisms by imagining a failed outcome. This can prompt useful questions, but the generated explanations are not established predictions. The cited 1989 research distinguishes temporal perspective from outcome uncertainty; it does not show that people are universally better at explaining than predicting.

It works on business plans, product launches, campaigns and investments, but just as well on career moves and personal goals: should I resign, should I go for this offer, is this investment sensible. The opening body also retains suggested trigger phrases: everyone on my team is enthusiastic and I find that suspicious, the plan sounds too good, we made exactly the same mistake last time, I do not dare to start and I do not know why.

One thing you should know beforehand: the skill is critical, not cynical. Every danger it names gets a countermeasure you can carry out this week, or an honest note that there is nothing to be done about it. Discouraging without giving you something to act on is literally on its list of refusals. The goal is not that you stop making plans, the goal is that your plans survive the inversion before reality carries it out.

Mitchell, Russo and Pennington studied temporal perspective and outcome uncertainty separately. Their 1989 abstract reports little temporal-perspective effect in the first experiment, while presenting outcomes as certain changed explanations and produced longer, more episodic accounts. This does not validate imagined causes as accurate predictions. The skill uses past-perfect wording as an exercise convention.

The second problem with the ordinary risk paragraph is one-sidedness: it may concentrate on money and miss other relevant categories. That is why the skill forces the mechanisms to spread across six categories, from demand and market to the silent assumptions nobody has tested. In assessment mode, where you have an existing plan checked, it even explicitly checks which categories are missing entirely, and checks people and assumptions without assuming they are always absent. Anyone who wants to see which thinking errors lie underneath, from optimism to groupthink, can turn to the Cognitive Bias Detector skill.

You see the difference fastest in diff form, with shortened illustrations based on the skill's hypothetical example:

Three times the same move: from a possibility with no address to an event with a mechanism inside it. A more specific mechanism can suggest a check or response, and that is exactly what the rest of the analysis does.

The seven actions begin with an agreed goal and review date, then an imagined failure question. The standard version generates eight to twelve mechanisms; the explicitly selected small-decision version has five. Both consider all six categories. Likelihood and impact use agreed 1-to-5 anchors; their product is an ordinal sorting aid, not a probability or expected-loss estimate.

A score of 12 or more OR late-only visibility marks a primary danger. Keep the visibility flag separately too. Every primary danger needs a feasible action or mitigation limit, a proposed owner/day to confirm, and an early-warning assessment. Other risks retain an action or monitoring status. The user keeps the decision.

The simplified score is inspired by FMEA, not a complete standard FMEA. The threshold 12 and separate visibility flag are local conventions. Each primary danger gets a possible warning or an explicit statement that no reliable early signal is known. Refusing to prepay does not prove future cancellation; the six-minute packing threshold is only illustrative.

There is also an assessment mode for existing work. Hand it a business plan, project plan or campaign brief, and you get an assumption scan with, per assumption, the question that makes it testable, the missing categories, an optimism check on figures with no source and growth that is simply extended in a straight line, ten failure mechanisms the plan itself does not mention, and the irreversible step: which step in the plan cannot be undone, and whether it comes early or late. Irreversible steps belong as late as possible. Anyone who wants to learn to build this kind of instruction file themselves finds the full explanation with template in writing a SKILL.md yourself NL.

The source list retains six entries, from Jacobi to Taleb and FMEA. Exact quotations, historic origins and predictions are not treated as verified simply because a name appears in that list. The download marks the unverified Jacobi wording and FMEA origin decade explicitly.

The evidence comes from decision research. Their 1989 abstract distinguishes temporal perspective from certainty: the first experiment found little temporal-perspective effect, while assumed certainty changed explanations. More detailed accounts are not necessarily accurate predictions. Gary Klein built the pre-mortem on that, described in Harvard Business Review in 2007: the team moves itself into a future in which the project has failed completely and writes down why. That technique has its own version in the library, the Pre-Mortem Analyst skill; the inverse thinker uses the same principle but builds the categories, the visibility flag and the countermeasures around it.

The scoring comes from industry. The source associates FMEA with defence and automotive history but does not establish the exact 1950s origin claim. This skill uses a simplified two-factor score and separate visibility flag, not a complete standard FMEA. The skill translates that to plans: a mechanism that only becomes visible in hindsight, such as customers who do not complain but simply cancel, counts more heavily than its score.

The variants come from Taleb and from entrepreneurial circles. Via negativa, worked out by Nassim Nicholas Taleb in Antifragile, is improving by removing: not asking what should be added, but what can come out without the plan collapsing. The skill uses that form when a plan is already overloaded, without claiming that overloaded plans are more common than underdeveloped ones. Anyone who wants to train Taleb's broader framework about systems that grow stronger from setbacks finds that in the Antifragile Thinker skill.

And the distinction between reversible and irreversible decisions, known in entrepreneurial circles as doors that swing both ways and doors that do not, determines how heavily the skill commits: a shortened inversion with five mechanisms for a decision you can undo tomorrow, the full analysis for the door that only swings one way. That distinction has its own skill, the Two Way Door Decider NL.

A simple exercise is to agree a goal and review horizon, imagine looking back on failure, and list specific possible mechanisms. The standard version uses eight to twelve, while a small reversible decision can use five. There is no evidence that the first three answers must be clichés or that the fourth becomes useful automatically.

After that, using it is simple: paste your plan and ask for the inversion. The more concrete your goal and your deadline, the sharper the mechanisms; whatever is missing, the skill asks about. New to skills? Read what Claude skills are NL first and the installation guide for Claude, Codex and Cursor NL. The broader explanation of working with AI is in the knowledge base.

If the plan is already overloaded, it switches to via negativa and asks not what should be added but what can come out. For a small or reversible decision, the shortened inversion with five mechanisms is enough, because a full analysis on a decision you can undo tomorrow is wasted effort. For investments above an amount that would actually hurt, a separate financial calculation gets added. And for personal decisions, such as resigning or moving house, the people category weighs more heavily and the skill explicitly asks what you would lose that cannot be expressed in money.

There are also situations where you are better off leaving it alone, and the file itself says so. If a decision has already been made and is irreversible, a failure list is nothing more than self-torment; the skill then shifts to limiting the damage. If you are mainly looking for confirmation or encouragement, this is the wrong skill, and that is not a dig: sometimes gathering courage is the real work, and a failure list does not help with that. And for legal, medical or financial questions it refers you on to people who carry professional liability.

The most honest limit is the same as with every thinking technique: the inversion does not make your plan good, it gives it questions and proposed tests. After the analysis you have mechanisms, alarms and countermeasures, but still no certainty. If you then want to work out what kind of decision you are actually facing and which approach fits it, the Cynefin Decider skill is the logical next step; anyone who, after the failure analysis, looks for calm in what lies outside their control finds the oldest form of negative visualisation in the Stoic Coach skill NL.

The failure question also works on marketing: how do we guarantee that customers never find us? Being unfindable on Google is then failure mechanism number one. Want to know whether that mechanism is already at play for you? The free SEO scan shows in a few seconds where your site stands. In the guide eight decision models compared NL you see how this model relates to the others. And if you want to talk further about what AI can do for your business beyond this, from separate skills to full automation, we simply do that in a conversation.