Claude Code Official Provincial Token Guide: 11 Tips to Make Credit Use Longer According to Beating monitoring, Anthropic has released the Claude Code provincial token guide. The core sentence is: the longer the context, the more expensive each subsequent round. Chatting, file and command output will continue to occupy tokens, which will also consume the quota faster as a result. In summary, there are 11 most practical tips: 1 Change the task/clean. Clear the context after completing a task. I am still working on the same task, but the previous content is useless. Please use/compress again. The longer the conversation, the more expensive it becomes, as each round will bring back the previous content. 2. Choose the model and effort from the beginning. Executing/model or/effort midway will crash the prompt cache, and the next round requires reprocessing the entire context. Fast mode is the same, it's best to start from the beginning. Before leaving, please coordinate. The cache of subscribed users expires in about 1 hour, and the API Key defaults to about 5 minutes. After the cache expires and returns, the next round usually involves reprocessing the entire context. Compressing while the cache is still available is cheaper. 4. In the past few rounds, I have been drifting or rewinding. It only deletes the later rounds, and the previous cache can still be used. /Compact will rewrite the entire conversation, which is more costly. 5. Use the @ file name directly for the file. Claude will receive the files directly in the first round, which can save one read or search. A file usually only needs @ once, and repeating @ may actually insert another copy of the same file into the context. 6. Prompt should be as specific as possible. Instead of saying 'test failed', it's better to directly tell Claude which test and which file. Otherwise, it may first grep, search, and read a bunch of files, and these results will continue to occupy context in each round. 7. Test and log outputs should be minimized as much as possible. Build、 The output of testing and Git logs will be left in context. You can add a quiet parameter, output only errors, or use tail truncation. Claude Code automatically writes commands exceeding 30000 characters to a file, leaving only the digest and path in the session. 8. Use/context to check the startup burden. At the beginning of the new conversation, you can see how much context CLAUDE.md, MCP, and tool definitions occupy. CLAUDE.md only includes general rules, while specialized processes are loaded into on-demand Skills. Unused MCPs are closed using/mcp. 9. Use smaller models and lower effort for simple tasks. The input and output of large models are both more expensive, and thinking also belongs to the output token. Pure execution tasks do not need to always use the highest level models and inference intensity. 10. Leave large logs to the sub agent, and do not misuse small tasks. Subagents have independent contexts and only bring back the final results to the main session, making them suitable for reading large amounts of logs and searching documents. But it also needs to reread files and run multiple rounds on its own, and small tasks may actually be more expensive. Tasks such as searching and reading logs can also specify Haiku or Sonnet. 11. It is best to open a separate session for/loop. Each iteration is a complete round that repeats the entire context of the current session. If the interval exceeds the cache time, it is possible to stack a cache failure. The official suggestion is to open a new terminal and use a clean session specifically for running loops. Anthropic's advice ultimately boils down to: less useless content, clear tasks promptly upon completion, and avoid prolonging conversations. [Original link]