← homeKatya Shalel
Free guide

Readable AI Decisions

AI already advises people on what to buy and whom to trust. Here's a system where every recommendation is explainable and verifiable by architecture, not by promise

Why it matters: a recommendation without grounds is a request to believe, and a user believes once. A system that shows the grounds of every decision wins not by honesty as a gesture, but by being checkable. Verifiability is the new currency of trust in a world where AI advises.

Step 1

Run the Indifference Test

Neutrality is checked with one question: does the system benefit when a specific product wins. If yes, the output is skewed regardless of intent. The Indifference Test walks the whole chain from monetization to ranking and finds every point where revenue is coupled to the outcome of choice.

Prompt · Indifference Test
Here's how monetization and ranking work in my recommendation system: [describe]. Run the Indifference Test: list every point where the system benefits when a specific product or brand wins. For each point, explain how it skews the output, and propose how to decouple revenue from the outcome of choice without killing the business model.
Step 2

A response schema that shows its grounds

A neutrality claim requires trust, a field in the API response does not. When every response carries the ranking basis, a sponsorship flag and the brand weighting, neutrality turns from a promise into a property of the architecture, verifiable by an outside auditor from a single response.

Prompt · verifiable response schema
Design the response schema of my recommendation API so that neutrality is verifiable inside the response itself. Required fields: ranking basis, sponsorship flag, brand weighting. Give one full example response and explain how an outside auditor would verify neutrality from that single response with no access to code or database.
Step 3

Explain the decision to the user

The schema is for the auditor, the explanation is for the human. Two or three lines: why this one, what it was chosen from, what would change the result. No superlatives, no marketing, only grounds. A user who sees the grounds comes back because the system can be checked, not because it demands belief.

Prompt · human explanation
Here's the system's structured response: [paste JSON]. Turn it into a two-to-three line explanation for the user: why this recommendation, what it was chosen from, what would change the result. Superlatives, marketing words and any claims not present in the response itself are forbidden.
Level up

Neutrality as architecture

Neutrality is not a footer statement, it's a property that lives in every response the system gives. In SKINBOT it's verifiable in the API response itself, the only format of a neutrality claim that requires no trust. Any system where AI advises and a human chooses is built on the same logic.

Go deeper

This is guide four of four, the last stage. The full course runs on a waitlist

First cohort · the waitlist closes August 6