945 runs later: /clear cuts Claude Code tokens 79%, high effort adds 20%
daniel_mac8 · x · 2026-10-18
Testing 12 popular practices across 945 Claude Code runs: /clear before tasks saved 79% tokens, /compact 60%, moving 16KB of rules from CLAUDE.md into a skill 31%, and flat-notes map with 'answer from the note' 32% (18/20 accuracy). Low effort on Opus 5.5 saved 17% at 20/20 correct. Worst: high effort +20%, Haiku 5.5 instead of Sonnet 5.5 +33% (though 12x cheaper per run), unrelated rules in CLAUDE.md +46%. No clear effect: disabling unused MCP servers (-8%), subagent for heavy steps (-3%), naming exact files +13% but 20/20 accuracy.
More from coding & agent
- Meta releases Muse Spark Open: 816B-parameter open model tops US open-weight rankings — arena · 2026-10-18
- 51 real-world AI agent setups analyzed: narrow tasks, drafts not sends, file-based handoffs — abe17124 · 2026-10-18
- User says $200 OpenAI plan drained in a day, then reset oddly to zero — alexcovo_eth · 2026-10-18
- Anthropic tip: one prompt sentence changes how Claude subagents pace speed vs thoroughness — yacinelearning · 2026-10-18
- wavoice: open-source CLI connects WhatsApp voice calls to LiveKit agents in 90s — yashubhakt · 2026-10-18
- "I want an agent to minimize my diffs" — even GitHub's web UI is obsolete now — davidcrawshaw · 2026-10-18