AI · 2h ago
Anthropic slashed Claude Code's prompt by 80% with no performance loss
Anthropic engineer Thariq Shihipar revealed the team removed over 80% of Claude Code's system prompt, shrinking it from 2,686 to 514 words, with no measurable drop in coding evaluations. The announcement garnered 4.3 million views in a week. Anthropic's audit found that contradictory instructions across the system prompt, CLAUDE.md, and skill definitions were making the model less reliable.
Meridian48 take
The finding that more instructions can degrade performance is a counterintuitive lesson for anyone building agent harnesses, but it's unclear how broadly this applies beyond Claude's latest models.
prompt-engineeringclaude-code