Excessive token usage: agent regenerated entire reports instead of surgical edits

What happened

During a session, I asked Claude to remove references to a specific username (me_sherif) from 3 dating report files. Instead of using Grep to find the ~10-15 lines containing the references and Edit to surgically remove them (~5k tokens), Claude launched 3 parallel subagents that each fully regenerated all 5 report files from scratch — reading all frameworks, analysis files, and schemas again.

Estimated token waste: ~350-400k tokens for a task that should have cost ~10-20k tokens.

Expected behavior

Claude should have:

1. Grep for "me_sherif" across the 3 report directories
2. Edit to remove/replace the specific lines
3. Done in <20k tokens

Actual behavior

Claude launched 3 full Task subagents (Sonnet model), each re-reading 10+ framework files and regenerating all outputs from scratch. Total: ~360k tokens, ~45 minutes.

Root cause

The agent defaulted to "regenerate everything" instead of "edit what's broken" — likely because the instruction mentioned reframing (cold outreach) alongside the removal request, and the agent over-scoped.

Environment

• Claude Code CLI
• Model: claude-opus-4-6
• Subagents: claude-sonnet-4-6