Token Efficiency Patterns

Concrete strategies for Claude Code to use fewer tokens while maintaining quality.

Tool Usage

• Parallel tool calls: Always batch independent reads, globs, and greps in a single message
• Glob before Read: Find the right file first instead of reading multiple wrong files
• Grep with head_limit: Use head_limit to avoid reading huge search results
• Avoid redundant reads: If you just wrote or edited a file, don’t re-read it to verify
• Use memory files: Check MEMORY.md before re-exploring the codebase for info you’ve seen before

Response Style

• Be concise: Short answers save tokens. Bullet points over paragraphs
• Don’t repeat the question: Jump straight to the answer
• Don’t explain what you’re about to do: Just do it
• Skip confirmation of obvious actions: “Done.” is enough
• Don’t list files you read: The user can see tool calls

Skills & Automation

• Use skills for recurring tasks: A skill call is cheaper than re-deriving a workflow
• Create skills proactively: If a task pattern repeats 3+ times, propose a skill
• Cache knowledge in vault: Write findings to 20_knowledge/ so future sessions don’t re-research

Codebase Navigation

• Use memory for key paths: Don’t re-explore project structure every session
• Targeted searches: Use specific glob patterns instead of broad **/* searches
• Read with offset/limit: For large files, read only the relevant section
• Use Explore agent for deep research: Keeps main context clean

Anti-Patterns to Avoid

• Reading a file, then reading it again in the same session
• Using Bash for cat, grep, find when dedicated tools exist (wastes tokens on tool overhead)
• Explaining every step before doing it
• Asking for confirmation on low-risk, reversible actions
• Re-reading CLAUDE.md or memory files mid-session (they’re already in context)
• Creating sub-agents for simple, single-query tasks