Why most people prompt AI incorrectly, according to Anthropic’s in-house philosopher

Artificial Intelligence

As AI chatbots spread into everyday life, the art of prompting has become a core skill. Yet there is no universal method that works across all models, and relying on standardised tips often leads to inconsistent results. Anthropic’s philosopher Amanda Askell says users need to rethink how they approach prompts if they want consistently reliable outputs.

In a new Q&A released by the company, Askell argues that prompting is not a rule-based discipline but an empirical one. There is no fixed manual. Instead, users must learn by experimenting, observing patterns and adapting to each model’s unique behaviour. She says that every new system requires a fresh approach, which only becomes clear through repeated interaction.

Askell urges users to drop assumptions about how a model “should” respond and instead study its actual outputs. According to her, good prompting demands patient examination of responses and careful refinements based on those observations. She adds that her philosophical training helps her express complex ideas clearly when communicating with models, which often improves the quality of responses.

Anthropic has previously offered its own framing for working with Claude. In guidance published in July, the company likened the chatbot to a brilliant but inexperienced employee with amnesia, one who needs highly explicit instructions each time. Because the model starts with no memory of a user’s preferences, norms or working style, clarity and context become essential.

The company emphasised that Claude performs significantly better when users provide detailed background, constraints and desired outcomes. The more context it has, the more precise and aligned its responses become. By treating each session as a fresh interaction and guiding the model with careful, explicit information, users can achieve far more reliable results.

In short, Askell’s message is clear: prompting is not a trick or a secret formula. It is a process of learning the language of each model through deliberate practice, clear expression and thoughtful iteration.