The AI Cognitive Bias Detector skill for Claude

The problem the skill solves sits in your own head, which is why you do not see it. Kahneman described two thinking systems: System 1 is fast, automatic and intuitive, System 2 is slow, deliberate and analytical. Most bad decisions are System 1 decisions presented as though System 2 made them. You feel that the price rise will work, and only afterwards does your analytical brain come up with the arguments for it. That is exactly what it reverses: System 2 gets a seat at the table before the decision is made, because the skill names the thinking errors explicitly, with evidence taken from your own wording.

It is meant for anyone making decisions that matter: an investment, a hire, a price change, a change in strategy, a move. According to its own description it also triggers on confident statements in absolute terms, such as this always works or everyone wants this, because the file says those almost always hide biases underneath. It is part of the skill library NL we make freely available from the AI and automation service, with no account and no sales email afterwards.

There is one thing you should know beforehand, and the skill says it itself in its closing lines: debiasing improves the process, not the outcome. A good process can still produce a bad outcome, and the other way round. What the scan does is reduce the chance that you make an important decision on reasoning you never actually checked.

System 2 is easily overruled. Deliberate thought costs energy, and your brain is frugal. Unless you actively switch System 2 on, it simply signs off on whatever System 1 had already decided. That is why a bad decision feels exactly the same from the inside as a good one: in both cases System 1 delivers a feeling of certainty that says nothing about the quality of the reasoning underneath it.

Spotting biases in someone else is easy, in yourself it is almost impossible. That is not a character flaw but the default setting of every human being, and the skill says so explicitly: including the user, and including the skill itself. That is exactly why an external scan works. Claude has no stake in your decision, fights no battle with your self-image, and reads only what is literally there.

Naming it alone is not enough. Its approach leans on the debiasing literature, from Larrick in 2004 to the training research by Morewedge and colleagues from 2015. The honest summary of that literature is stated plainly in the file: naming a bias does not remove it. What does help are structural interventions, such as looking up base rates, setting a decision rule in advance and organising contradiction. That is why every scan ends, without exception, in an action list rather than a diagnosis.

Here is what that looks like at sentence level, using lines from the skill's own example: the confident System 1 claim out, the testable System 2 question in.

Then comes the mandatory six-phase process, and the file is strict about it: skip no phase. Phase 1 is the intake, in which Claude reads the input literally without summarising, and notes what the core claim is, which assumptions are being made, what emotional charge sits in the wording, and which alternatives are not being considered.

Phase 2 is the diagnosis: is this predominantly System 1 or System 2, expressed as a score from 1 to 10 with a justification of at most three sentences. Phase 3 is the actual bias scan across the six families from FIG.02, giving each spotted bias its name, a one-sentence definition, the quote from your input as evidence, and an impact estimate from low to high.

Phase 4 ranks the three biases that distort the judgement most and describes concretely what the most likely wrong outcome is if they remain active. Phase 5 delivers a System 2 reframe per top bias, and the file requires it to have three parts: an explicit counter-question, an alternative scenario and a measurable check. Phase 6 closes with a debias action list of three to five actions for within 24 hours, using techniques named with their source in the file: the consider the opposite exercise from Lord, Lepper and Preston in 1984, the pre-mortem from Gary Klein in 2007, looking up base rates, a red team conversation, and parking the decision for 24 hours.

The file also fixes the answer format, right down to the headings; you can see that below in FIG.03. And it contains seven important rules on tone and evidence, including the two that keep the scan honest: no quote, no bias, and if no biases seem to be active, say so, because false positives are biases too. If you want to learn how to build an instruction file like this yourself, the full explanation is in writing a SKILL.md NL.

The debiasing techniques in phase 6 each have their own source: consider the opposite comes from Lord, Lepper and Preston in 1984, the pre-mortem from Gary Klein in 2007, and the overview of what does and does not work from Larrick 2004 and Morewedge 2015. The distinction between bias and noise, two different problems that often get lumped together, comes from Noise by Kahneman, Sibony and Sunstein from 2021.

More striking is what comes after that: a chapter on the limits of its own framework, with the instruction to be open about them if you probe further. Bias labels are interpretations, not measurements; two analysts can classify the same sentence differently, which is why the quote matters more than the name of the bias. The System 1 score is a conversation opener, not a validated scale. And the file states explicitly that part of the classic bias literature has held up more weakly in replication research than the original publications suggested: treat bias names as thinking tools, not laws of nature.

Also honest: biases are not always wrong. Fast heuristics work excellently in stable environments with a lot of repetition, the file writes. They go wrong on rare, large and hard-to-reverse decisions, and that is what this skill is for. Anyone who wants to turn that distinction between reversible and irreversible into a decision method will also find a separate WRAP Decider skill NL in the library, and for quantifying uncertainty with base rates and priors there is the Bayesian Thinking Coach skill.

Even without installing the skill you can improve your own decision with these principles: write down your prediction before you decide, look up the base rate before you trust your own estimate, and ask yourself what you would need to see to change your mind. If you cannot answer that last one, according to the skill you do not have a decision, you have a wish.

After that, using it is simple: paste your decision, plan or belief into the conversation and ask for a bias check. The more literally you write down your own reasoning, the better the scan, because the quotes are the evidence. If your input is too thin, the skill, by its own rules, asks exactly one clarifying question and then carries straight on. If you get stuck anywhere during installation, the full step-by-step story per environment is in installing Claude skills NL, and the broader explanation of working with AI is in the knowledge base.

There are also situations where you are better off leaving it alone, and the file names them itself. Routine decisions in a stable environment with a lot of repetition: your fast heuristics work excellently there, and scanning every small choice mostly just makes you slow. Decisions that need subject-matter advice: the skill gives no medical, psychological, legal or investment advice, as a matter of principle. And thin input: with half a sentence of context, the skill, by its own rules, says the analysis is provisional, because without context every scan is weak.

The most important limit is the same one that applies to any form of debiasing: the scan improves your process, not your outcome. If the offer is not sound or the market turns against you, clean reasoning will not save you. What the skill does prevent is discovering an avoidable thinking error only once the cancellations start coming in. And if the thinking error is not in the decision itself but in the assumptions you are stacking it on, look at the Ladder of Inference Coach skill, which walks precisely that path from observation to conclusion.

A debiased decision about your marketing is nice, but most businesses have a different problem first: they do not know where they stand. The free SEO scan shows in a few seconds how findable your site is right now, with data instead of gut feeling, which fits exactly in the spirit of this skill. And if you want to talk through what else AI can do for your business, from stand-alone skills to full automation, we simply do that in a conversation.