cost-aware-session

Running Claude Code Cost Aware

Most cost in a long session comes from a few repeatable patterns: reading more than you need, redundant verification, broad searches that dump raw output into context, and reaching for a heavier model than the task needs. This skill is a checklist for catching those, not a one time setting.

Steps

1. Match the model to the task.

   • Routine, well scoped work (drafting a standup note, formatting, simple lookups, running a known command) doesn't need your most capable model. Lighter/faster models handle these well and cost less.
   • Reserve deeper reasoning models for genuinely hard problems: ambiguous debugging, architecture trade offs, multi step planning.
   • .claude/settings.json pins a sensible default; only override per task when the task's difficulty actually warrants it.

2. Read narrowly.

   • Grep for the symbol or pattern first; read only the file and line range that matters. Reading whole large files "to be safe" burns tokens for content you'll never use.
   • Don't re read a file immediately after editing it. A successful edit means the change landed; the harness already tracks that state.

3. Search with scoped subagents instead of one sprawling search.

   • A single broad search across many repos pulls all of that raw output into your main context. Dispatching several narrow, read only agents in parallel (see triage-researcher) keeps each one's output compact and your own context clean, which is cheaper over the whole session even though it looks like "more agents."

4. Batch and schedule instead of repeating.

   • If you find yourself running the same broad query every morning, that's a candidate for a scheduled routine that filters for what's actually notable and skips posting when there's nothing to report (see jira-progress-pulse). One filtered scheduled run costs less over a month than many manual broad ones.