Claude: System Prompts

	Claude: System Prompts(platform.claude.com)
	743 points by tosh 58 days ago \| 280 comments
	tl;dr: Summary not available
	HN Discussion: •Providing supplementary resources like git history and full prompt extractions for tracking changes ↓System prompts are too long and noisy, hurting model performance versus shorter guidance ↓Enforcing basic common sense via system prompts suggests Anthropic doesn't fully trust model intelligence ↓Heavy system prompts make models unhelpful; users prefer raw models or modular prompt choices ↑System prompts reveal interesting behavioral roadmap and safety layering strategy