Anthropic shares 6 official tips to cut Claude Code token costs
addyosmani · x · 2026-09-03
Anthropic engineer Lydia Hallie published an official guide on maximizing Claude Code session efficiency:
- Run /clear between tasks to avoid resending irrelevant context and cut token usage.
- Set model and effort level upfront — switching mid-conversation busts the prompt cache and raises costs.
- @-mention files instead of naming them, attaching files directly and saving a Read call or search.
- Silence noisy command output or run it in a subagent, since output lingers in the session like a file.
- Run /context once in a fresh session to see what's loaded (CLAUDE.md, MCP tools) and trim the excess.
- /compact before stepping away — the prompt cache expires after an hour, and summarizing is cheapest while still cached.
With agentic coding tools billed by usage, every task now has a real cost, making context management the key to savings.
More from coding & agent
- OpenClaw isn't dead: cheap Chinese models may push personal agents mainstream — FuSheng_0306 · 2026-09-03
- The /show-me skill makes PR descriptions dramatically easier to read — mattpocockuk · 2026-09-03
- ROS 2 gotcha: use node.get_clock().now(), not time.time(), in simulation — 4310sy · 2026-09-03
- Matt Pocock praise show-me skill: pseudocode and tree diagrams that make PRs readable — mattpocockuk · 2026-09-03
- Opus 5 all-day usage report: no parallel subagents means limits are a non-issue — 4310sy · 2026-09-03
- The easiest setup to access your local AI from anywhere via phone — TheZachMueller · 2026-09-03