Post this publicly (GitHub gist / repo README / Reddit / Show HN). It's the runnable, free-to-try hook that makes Show HN and Reddit value-first posts legitimate. The full kit (18 tactics + drop-in configs + rules + scripts) is the paid upgrade.
Route the model to the task. Haiku for mechanical edits/scaffolding, Sonnet for daily feature work, Opus only for genuinely hard reasoning. Pinning everything to the top model is the most expensive default there is.Every turn re-sends the whole conversation as input tokens; don't pay to drag a finished task into the next one./clear
between unrelated tasks.Stop re-reading files you just edited. The edit already confirmed success — re-reading to "verify" doubles the cost for zero new info.(API) Turn on prompt caching + the Batch API. Cached input and non-realtime batch jobs are dramatically cheaper. Most people never switch them on.Cap extended thinking on routine sessions. Great for hard problems, pure overhead on easy ones.
A read-only script (Windows PowerShell + bash) that scans your local Claude Code transcripts and shows input/output/cached tokens + cache-hit % over the last N days. Run it weekly; watch the number drop. (Included in this sample.)
The paid Claude Code Token-Saver Kit adds:
- The full 18-tactic playbook, ranked by $ saved per minute
- A ready-to-merge cost-control
settings.json
(model default, thinking cap, hygiene hook) - A model-routing decision table
- A drop-in
cost-control.md
rules file - Single-seat commercial license + 14-day refund Prices and model names change — always confirm current numbers in Anthropic's official docs. The tactics are durable.