token-optimization

Run a session audit with /context and /plugin, disable unused MCPs, add concise-response directives to CLAUDE.md, and filter verbose CLI output with RTK. Documented results show reductions from $750/month to $100/month.

Adding MCPs or tools mid-session, modifying CLAUDE.md, or editing earlier messages invalidates the cache and re-bills affected tokens. Configure everything before starting, and use /compact plus a new session if changes are required.

Spawn agents with context forking so they work in isolated windows and return only results, not full history. Spawn them early while the parent context is still light, and delegate web research and transcription tasks to subagents.

RTK is a CLI output filter installed via npx rtk init --global that strips verbose command output before it enters context. It reduces commands like git status from roughly 50 lines to about 5, saving around 80-90% on those calls.

Only for very large codebases. For small to medium projects, set contextWindow to around 200,000 in settings.json, since a constrained window enforces better context discipline and reduces bloat.

Common causes are too many loaded MCPs, screenshot-based browser automation, verbose command output, and missing context forking for agents. Run /context to identify the source, then apply the skill's phased optimization steps.