Zack Witten

Zack Witten is an AI researcher and prompt engineer known for making language-model behavior more reliable while investigating hallucinations, evaluation failures, and the peculiar preferences models sometimes exhibit. At Anthropic, he progressed from helping individual customers solve prompting problems to developing broader prompting tools and educational resources for Claude.

From computational biology to Claude

Earlier, Witten collaborated with his brother Jacob on the GRAMPA antimicrobial-peptide dataset, using neural networks to predict whether peptide sequences could inhibit bacteria. Their open-source antimicrobial-peptide project combines experimental measurements with models targeting organisms including E. coli and P. aeruginosa. His analysis of train-test splitting also showed how ignoring relationships between observations can inflate apparent machine-learning performance.

At Anthropic, Witten translated hands-on customer work into reusable prompting guidance and tools, including a prompt generator. He described that transition in an Anthropic conversation about prompt engineering, led a live Claude prompting workshop at AI Engineer World’s Fair 2024, and contributed feedback on Anthropic’s January 2026 constitution for Claude.

• Use deterministic code before clever prompting. Witten favors assistant prefills, output parsing, stop sequences, and programmatic validation for structured JSON generation. When ordinary software can enforce a requirement, he avoids assigning it to an unpredictable model.

• Examples define the task. Carefully selected positive and negative examples communicate subjective requirements more effectively than vague instructions. He prefers a few representative, high-quality examples, tested against concrete cases.

• Ground answers before generating them. For hallucination-resistant prompting, Witten has models identify supporting material before drawing conclusions. He distinguishes useful reasoning performed in advance from retrospective explanations, which can rationalize an answer after the fact.

• Treat model preferences as testable behavior. His experiments with model-specific interests compared how Claude models allocated hypothetical conversations across subjects. A tic-tac-toe experiment explored whether models changed their play when a user appeared distressed, without assuming humanlike motives.

• Hallucinations can contaminate the web. Witten traced Gustav, a fictional hippopotamus, from an invented prompting example into generated web content and search results, showing how repeated fabrication can resemble independent corroboration.