Anthropic

What does it take to be an AI whisperer?

What does it take to truly understand how AI models think? Anthropic researcher Amanda Askell shares what it means to be an “LLM whisperer.”

AAAmanda Askellguest

Dec 15, 2025·0m·Watch on YouTube ↗

📖 CHAPTERS

1. 1

   What it means to be an “LLM whisperer” at Anthropic

   The clip frames the core question: what skills and habits make someone effective at eliciting good behavior from large language models. It sets the context that “whispering” is less about mystique and more about hands-on practice.

2. 2

   High-volume model interaction to build intuition

   Amanda emphasizes that a key ingredient is spending lots of time interacting with models and closely inspecting many outputs. This repetition helps develop an intuitive sense of the model’s “shape” and response tendencies.

3. 3

   Experimentation mindset: try, observe, iterate

   The work is described as inherently experimental—trying variations, observing results, and iterating. Progress comes from a willingness to test hypotheses about prompting and model behavior.

4. 4

   Prompting as an empirical discipline (not just “prompt tricks”)

   Amanda highlights that the domain is surprisingly empirical, which she suggests is often misunderstood. Success relies on evidence from actual model behavior rather than assumptions about how it “should” work.

5. 5

   Communicating concerns clearly to the model

   A practical tactic she uses is to explain an issue or concern to the model as clearly as possible. Clarity in the user’s intent is positioned as foundational for getting reliable, aligned responses.

6. 6

   Diagnosing unexpected outputs: ask “why” and trace misunderstandings

   When the model does something unexpected, Amanda suggests two complementary strategies: asking it why it responded that way, or analyzing what in the prompt led to misunderstanding. This is presented as a core debugging loop for model interaction.

7. 7

   Why the work is compelling: discovering depth in model behavior

   She closes by noting how interesting the work is and how it reveals surprising depth in the models. The job isn’t just about getting correct outputs—it’s about exploring and understanding complex behaviors.