Update your confidence, not your story, with a Bayesian thinking skill

How it works

A fixed route keeps the work reviewable. Each step produces input for the next one.

Why a prior belongs on the page

People update beliefs all the time, but usually hide the starting point. A new lead feels like proof that a market exists; a lost deal feels like proof that it does not. Without a stated prior, both reactions can be defended after the event. The skill asks what you believed before the signal and what evidence created that belief.

A prior does not need a decimal. High, medium or low confidence with a reason is often more honest than 63 percent. When credible base-rate data exists, Claude can use it. When it does not, the skill keeps the estimate qualitative and names what would make it stronger.

Evidence is not counted, it is weighed

Five articles repeating one press release are not five independent signals. A customer interview is relevant to that customer but may not describe the market. A conversion result from a tiny sample can be promising without being stable. The skill separates source reliability, directness, independence and sample quality before it updates anything.

It also asks what else could explain the observation. A campaign may improve because demand rose, tracking changed or the offer became cheaper. Naming alternatives does not cancel the result. It stops one convenient explanation from becoming the only explanation.

you > /ai-bayesian-thinking-coach
claude > intake → method → checked output
[ok] assumptions remain visible

What the update looks like

The answer shows the original claim, prior, strongest supporting evidence, strongest opposing evidence and the resulting confidence. It explains the movement in words. If you supplied real probabilities and conditional evidence, it can show the calculation, but it never manufactures a likelihood just to make the answer look scientific.

The final line is practical: keep the decision, change it, pause it or run a test. Claude also proposes the next observation with the highest information value. The point is not to sound uncertain forever. It is to spend uncertainty on the right experiment.

Use it for decisions that can still move

The skill works well for product demand, marketing signals, hiring assumptions, project risks and forecasts that will receive more evidence. Give it the claim, the starting view and the new information. If different people hold different priors, record them separately before discussing the update; that disagreement is useful evidence about the decision process.

It is not a substitute for a regulated professional, a statistical model or a risk system. Medical, legal and financial decisions need qualified review and validated data. The skill can organise questions for that review, not replace it.

A small example without false precision

Suppose three prospective customers ask for the same feature. The prior is that the feature may help a narrow segment but is not yet a market requirement. The signal supports demand, but all three leads came through one partner and may not be independent. Confidence moves upward a little, not to certainty.

The useful next step is not a full build. It is a short prototype shown to customers outside that partner channel, with a predefined decision rule. The update has changed the action while keeping the uncertainty visible. That is Bayesian thinking in ordinary work.

Continue from here