← back
Principles for Prompt Engineering - Karina Nguyen (Claude Instant @ Anthropic)
Takeaway
Prompting Claude well means treating it like a person, structuring inputs with consistent XML tags, and iterating like creative writing.
Summary
- Karina Nguyen (Anthropic, worked on Claude self-correction and hallucination reduction) frames prompting as creative writing — hypothesis-driven iteration, not luck.
- Three reasons prompting fails: humans know what they want but not how to ask, vaguely know, or don't know at all; few-shot examples fix the middle case.
- Six writing principles: Clarity, Conciseness, Coherence (context first, task last), Consistency, Direction (style/length guidance), Grounding (quote sources).
- Claude was originally fine-tuned on XML tags — using <user_query>, <answer> etc. consistently materially improves Claude performance over other formats.
- Demos Claude as a recommendation system over a CLIP image-text database with zero-shot relevance scoring in answer tags.
prompt-engineeringclaudeanthropic
Original description
Learn the state of the art on prompt engineering from experience leading engineering on Anthropic Claude.