Claude Code cuts claude-api Skill context
Anthropic reduced Claude Code’s built-in claude-api Skill from more than 200,000 tokens to roughly 25,000 by applying progressive disclosure. The redesign loads detailed API knowledge only when relevant, preserving context for the actual coding task.
This is a strong reminder that agent quality depends as much on context architecture as model capability.
- –Smaller always-available instructions leave more room for code, tool output, and reasoning
- –Progressive disclosure turns large skills into navigable knowledge bases instead of prompt-sized manuals
- –Skill maintenance should be treated like software maintenance: prune duplication, split references, and test retrieval paths
- –The approach scales better as teams add specialized skills, agents, and project knowledge
DISCOVERED
2h ago
2026-08-19
PUBLISHED
2h ago
2026-08-19
RELEVANCE
AUTHOR
dani_avila7