6 articles in this category
6 articles
Systima.ai research reveals Claude Code sends 33,000 pre-prompt tokens per initial request, consuming 4.7 times more baseline context than OpenCode.
Builder LabA developer's blog post, corroborated on Hacker News, found that Claude Code embeds invisible Unicode characters in generated code — Anthropic hasn't confirmed why, but the characters could break dev
Builder LabAnthropic resolved three separate issues in Claude Code that degraded model output, including a prompt change that cut eval scores by 3%.
Builder LabAnthropic introduces Claude Code Routines in research preview, allowing developers to automate multi-step coding workflows across cloud environments.
Builder LabA GitHub issue by developer stellaraccident analyzes 234,760 Claude Code tool calls, linking reduced thinking-token budgets to a reported 70% drop in pre-edit code research.
Builder LabOpen-source MCP proxy Context Mode compresses Claude Code tool outputs by up to 98%, cutting 315 KB of raw data to 5.4 KB using local subprocess sandboxes.
Builder Lab