Claude Code Prompt Shrinks as Models Improve
Anthropic removed over 80% of Claude Code’s system prompt for Claude Opus 5 and Fable 5 without measurable coding-evaluation losses. The team says stronger models need fewer behavioral patches, shifting agent design toward judgment, progressive disclosure, and verification.
The real lesson is that prompt accumulation can become technical debt: scaffolding built for weaker models may constrain stronger ones.
- –Audit system prompts, CLAUDE.md files, skills, and hooks after every major model upgrade
- –Replace rigid rules with context-sensitive guidance that lets the model exercise judgment
- –Use progressive disclosure so specialized instructions load only when needed
- –Move reliability-critical behavior into tests, hooks, permissions, and tool interfaces
- –Treat prompt deletion as an ablation experiment, restoring instructions only when failures justify them
DISCOVERED
1h ago
2026-08-14
PUBLISHED
1h ago
2026-08-14
RELEVANCE
AUTHOR
olihels