Taste as Predictive Models

Taste is the quality of your predictive models and search heuristics: if I design the experiment this way, what will I find? If I write the doc this way, how much will it resonate? If I build the tool this way, how easy will it be to use?

Good taste = accurate predictions about what will work, arrived at quickly.

How to find where your taste is best

High-quality taste is highly idiosyncratic — most people's best taste is concentrated in specific, sometimes narrow domains. The diagnostic question:

What does it seem like everyone else is mysteriously bad at?

That persistent sense that "this is obviously wrong and nobody else seems to notice" is a signal that you have good taste in that domain. Don't wait for this to feel like confidence or competence — it usually feels more like mild frustration with others' blind spots.

How to improve taste

Think explicitly about predictions. When deciding what to do, ask "what do I predict will happen if I choose option A?" and try to unroll the trajectory. Even when you think you're already doing this intuitively, being explicit reveals things that intuition misses.
Revisit your predictions. After a decision resolves, check how your prediction held up. The feedback loop is what actually builds the model.
Metacognition. After projects, ask: what should we have done differently? What would we have needed to know earlier? What should I weight differently next time? Each lesson is small individually but compounds over time.

The most effective people at Anthropic do the most metacognition — they're also the most likely to identify improvements to processes and mental models.

What taste is not

Taste is not smartness or confidence. It's not evenly distributed across domains. Someone with excellent taste in experiment design may have poor taste in what blog post titles get upvoted. The useful question is always: where is your taste good?

Sources

impact-agency-and-taste-ben-kuhn