Cost Tracking

When to Use Subagents

• Exploratory research that may consume many tokens
• Tasks that don't need the full conversation context
• Parallel independent subtasks

When to Stop and Re-Evaluate

• Same error has been attempted 3+ times (token waste)
• Exploring broad codebase without narrowing (use targeted search instead)
• Generating large amounts of code that will need to be rewritten

Relative Model Costs (GitHub Copilot Pro+)

See skills/model-routing.md for the full decision framework.

Token-Efficient Patterns

1. Targeted reads — read specific files/sections, not entire directories
2. Grep before read — find the right file first, then read it
3. Batch operations — make multiple independent tool calls in parallel
4. Avoid re-reading — reference files already in context instead of reading again
5. Concise output — shorter responses use fewer output tokens

Anti-Patterns

• Reading every file in a directory "just in case"
• Repeatedly reading the same file across multiple tool calls
• Using premium models for simple tasks
• Generating verbose explanations when the user asked for code
• Re-exploring codebase context that was already gathered

See Also

• model-routing — Model selection with cost awareness