The AI First Principles Thinker skill for Claude

The contrast that makes the method so useful sits between two ways of reasoning. The first is reasoning by analogy: you look at what others do and adapt it. Fast, safe, and it never delivers more than a variation on what already exists. The second is reasoning from first principles: you break the problem down to what is demonstrably true and build back up from there. Slow, uncomfortable, and the only route to an answer that departs from the norm. That is exactly how the skill file puts it too, and that phrasing is sharper than the usual description because it states the price of the method straight away.

The AI First Principles Thinker translates that way of thinking into a fixed method for Claude. A skill is an instruction file, SKILL.md, that gives an AI assistant a fixed approach for a task. No software, no subscription, no connection to your systems: a text file of 2,744 words that tells Claude how to strip a plan down, how many assumptions to surface, how to score them, what output to deliver, and what it must never make up under any circumstances. You download the zip at the top of this page, put it in your Claude environment, and from that moment every plan you put forward goes through this route.

What sets the skill apart from just thinking critically along with you comes down to three things. First, when it breaks a plan down it separates three kinds of statement: laws of nature and mathematical truths that never change, contractual and legal facts that can change but not because you want them to, and conventions. That third category is where the gain lies, because conventions present themselves as facts. Second, it writes every assumption so that it could turn out to be wrong. Not that the market is big enough, but a claim with a number and an amount in it that you can check. Third, it keeps probability and impact strictly apart, each as its own figure from 1 to 5, because a merged gut-feel score does not point you anywhere.

This file belongs to our skill library NL, where 100 skills are free to download, and comes out of our AI and automation service. If you are working with this kind of file for the first time, start with what Claude skills are NL: it explains in plain language what a skill can and cannot do.

The convention that passes itself off as a law of nature. That is how we do it here, that is how the market works, that is how it is supposed to be. This is the richest layer, because conventions are invisible on purpose: they only work as long as nobody notices them any more. The breakdown block in the output forces the separation by putting three lists one below the other, and the file adds, drily, that the third list is usually the longest and the most interesting. In the case study that comes out as something like: that a course has to cost €495 because others charge that. Put that way, it is no longer pricing policy but an assumption with a price tag on it.

The claim that could never turn out to be wrong. The market is big enough. Our target audience needs this. Sentences like that feel like analysis but are not, because there is no conceivable outcome that would disprove them. The skill's writing rules forbid them and immediately give the alternative: write a claim with a number and an amount in it instead, so that you can check it. In review mode that is a separate check, the falsifiability check, which rewrites every unfalsifiable assumption into a claim that can be tested.

The judgement that folds probability and impact into one. This is a big risk: sounds like a score, points you nowhere. You do not know whether it is big because it is likely to go wrong or because it is expensive if it does, even though you act very differently on those two. With a high probability and low impact you run a quick, cheap test. With a low probability and high impact you build a safety net. That is why there are always two separate figures, and merging them into a gut-feel score is explicitly on the list of things it never does.

There is a fourth thing underneath all three, and it is stated right in the opening lines of the file: the skill confirms nothing out of politeness. That is not a matter of tone but a working instruction. The assumption the whole plan rests on, and that nobody has checked, is usually exactly the assumption everyone in the industry shares, and it survives every conversation in which the participants want to be nice to each other.

In diff form, with the same claims first as they usually appear in a plan and then as the skill rewrites them:

It opens with frontmatter stating when Claude should pick up the skill. Not just for the explicit terms like first principles thinking, first principles, thinking from the ground up, surfacing assumptions, assumption mapping, riskiest assumption and breaking down a cost structure. Also for questions that have nothing to do with a framework: does my business plan hold up, what am I missing, why is this so expensive, can this be fundamentally different.

And for exasperated remarks you would sooner hear in a canteen than in a strategy session: everyone in the industry works this way so we do too, we copied this from a competitor, my investor asks why this works and I do not have a good answer, our price is based on what everyone else charges.

Right below that sits the role description, and it is unusually strict for an AI instruction: the skill confirms nothing out of politeness, hunts for the assumption the whole plan rests on that nobody has checked because everyone in the industry shares it, and never makes up figures, market data or research findings.

The core is a list of eight fixed actions. It rewrites the goal into a measurable sentence with a deadline and strips out the solution if one is already hidden inside it. It strips the plan down and separates laws of nature and mathematical facts from contractual facts and from conventions. It surfaces eight to twelve assumptions, spread across the customer, the market, execution, technology, money and regulation, each as a falsifiable claim. It scores every assumption on probability and impact, from 1 to 5, and records the evidence strength.

After that it gets sharper. It flags the highest-scoring ones as primary risks, usually two to four. For each risk it delivers a redesign step you can carry out this week plus the cheapest test. It rebuilds the plan from what remained and states for each step whether it departs from the original plan and why. And it closes with the risks you consciously accept, so that there is a choice in place of a blind spot.

Before anything happens, the skill runs through a mandatory input checklist of seven points: the plan in your own words, the deadline by which it has to succeed, what is at stake in money, time, reputation or business, where the plan came from, the hard constraints, what you already know for certain and what that is based on, and what has already been tried in this direction. It only asks about what is missing and then stops asking, so you do not get a form, just a handful of targeted questions at most.

The output is fixed into eight blocks in a set order: the goal sharpened, broken down into three lists, the assumptions, the score table, the primary risks, the redesign per risk, the plan rebuilt in five to eight steps, and consciously accepted. The file also contains a worked case study about an online course, six situations where you pick which variant, a review mode with six checks for an existing plan or an existing risk analysis, seven things the skill never does, and six sources. If you want to see how a file like this is technically put together, from frontmatter to refusal list, writing a SKILL.md NL explains that step by step.

A striking number of rules are about language, and they are not all cosmetic. Clear Dutch, addressed as you, no dashes in running text, AI in capitals, no lists of exactly three, concrete over abstract, a figure over an adjective. But the last writing rule is substantive and maybe the most important in the whole file: word every assumption so that it could turn out to be wrong. That one rule decides whether the rest of the analysis is about anything at all.

The idea comes from Aristotle. He wrote about first principles: the starting points you cannot derive from anything else and on which all further reasoning rests, among other places in the Metaphysics and the Posterior Analytics. The term first principles traces back to that work. That is exactly what the breakdown block does. Not searching for the best explanation, but for the layer that needs no explanation at all.

The method comes from Descartes. His Discours de la méthode from 1637 made methodical doubt the road to certain knowledge: set aside everything that can be doubted and only then build back up. The two-part break-down-and-build-up you see in FIG.01 and FIG.03 comes from here.

The scoring comes from Bland and Osterwalder. Testing Business Ideas from 2019 worked out mapping and testing assumptions for entrepreneurs, in the form of an assumption map that plots importance against the amount of evidence. That is where the idea comes from that an assumption has two independent properties and that you must not fold them together. If you then want to actually run and measure those tests, the Mom Test Coach skill connects directly to that: it is about customer conversations in which people do not just tell you what you want to hear.

The premortem comes from Gary Klein. His article Performing a Project Premortem in Harvard Business Review from 2007 describes the exercise in which you project yourself forward to a moment when the plan has failed and write down the story of that failure. It is named as a variant and comes with a practical rule for teams: have everyone write separately before you bring them together, otherwise everyone follows the first speaker. As a stand-alone skill, that exercise lives at the Pre-Mortem Analyst NL.

Inversion comes from Jacobi via Munger. Charlie Munger popularised inversion as a thinking tool, collected in Poor Charlie's Almanack from 2005, and the rule itself is generally attributed to the mathematician Carl Jacobi. You do not ask how this succeeds but how this is guaranteed to fail, and then avoid that route. That one also has its own file: the Inversion Thinker.

Its business fame comes from Elon Musk. He used the approach to reduce the cost structure of rockets and batteries down to the price of the raw materials, instead of starting from what such parts had historically cost. The file is careful about this: that application became widely known through public interviews and cannot be attributed to a publication. In the skill the same move is called the cost breakdown, and you can turn it loose on any price by breaking it into raw material, labour, transport and margin, and asking of each component why it is what it is and who sets that.

Even without installing the skill, you can get further with these principles: split your plan into three piles, rewrite every claim into something that could turn out to be wrong, give probability and impact two separate figures, and for the top two work out the cheapest way to find out.

After that, using it is simple: describe your plan in your own words, with the deadline attached and what is at stake. You do not need to phrase it neatly, because it is precisely the messy phrasing that gives away the conventions. It runs through the input checklist, only asks about what is missing, and then works through its eight steps. If the installation gets stuck anywhere, the full explanation per environment is at installing Claude skills NL, including the mistakes everyone makes with folders and permissions. And if you want to learn to work more broadly with AI, you will find the background articles in the knowledge base.

If a team is involved, the premortem works better than a conversation, with a practical rule attached that saves a lot of sessions: have everyone write the failure story separately before you bring them together, otherwise everyone follows the first speaker. And if a single claim carries the whole argument, socratic questioning is enough: keep asking how you know that, until you land on evidence or on thin air.

There are also situations where you are better off leaving it alone, and unlike many skills, that is stated in the file itself. Skip it for low-cost routine jobs, because thinking from the ground up costs time, and you only earn that time back on decisions that are hard to reverse. And do not use it if you just want a decision confirmed, because the skill confirms nothing out of politeness. That is not a minor caveat: if the decision has effectively already been made and the analysis is only there to justify it, this skill mainly produces friction.

Besides drafting, it can also assess. Hand it an existing plan, a business case or a risk analysis, and you get six checks back. The convention check flags every sentence that presents a choice as a given, with sentences starting with that is just how it is or in our industry as prime candidates. The analogy check names what has been copied from someone else and what the evidence is that it works there.

The other four checks are about the assumptions themselves. The falsifiability check rewrites every assumption that could not turn out to be wrong. The completeness check looks at which category is missing, and the file predicts which two that usually are: regulation and the assumptions about the customer. The scoring check looks at whether probability and impact were scored separately or folded into a judgement that points nowhere. And after that comes a rebuilt version following the standard structure.

One more boundary worth naming, and one the file does not state: the breakdown produces clarity, not certainty. After the score table you know which two to four claims your plan rests on, and that is progress, because before you did not know that either, you just thought you did. But you still have to test those claims afterwards, and that costs time or money. Anyone who uses the outcome to tear everything up without running a single test has turned the method on its head. If you want to go from a problem that has already occurred down to the cause underneath it, the 5 Whys Analyst skill is the tool for that, not this one.

If one of your assumptions is about visibility, for instance the assumption that customers will find you anyway, that one is easy to test. The free SEO scan shows within seconds where your site stands, and that is evidence instead of a feeling. And if you want to talk through what else AI can do for your business, from individual skills to full automation, we simply do that in a conversation.