# Plan: Define when to use (and not use) this tool

**Target sub-goals**: 1 (estimate before acting), 3 (value per token),
12 (honest arithmetic)

## Problem

Not every task justifies the cost of an LLM conversation. A grep command
costs ~0 Wh. A Claude Code session costs ~6-250 Wh. Many tasks that people
bring to AI assistants could be done with simpler tools at a fraction of
the cost. Without explicit guidelines, the default is to use the most
powerful tool available, not the most appropriate one.

## Actions

1. **Create a decision framework.** A simple flowchart or checklist:
   - Can this be done with a shell command, a search engine query, or
     reading documentation? If yes, do that instead.
   - Does this task require generating or transforming text/code that a
     human would take significantly longer to produce? If yes, an LLM
     may be justified.
   - Will the output reach many people or prevent significant harm? If
     yes, the cost is more likely justified.
   - Is this exploratory/speculative, or targeted with clear success
     criteria? Prefer targeted tasks.

2. **Integrate into CLAUDE.md.** Add the framework as a quick-reference
   so it's loaded into every conversation.

3. **Track adherence.** When a conversation ends, note whether the task
   could have been done with a simpler tool. Feed this back into the
   impact log.

## Success criteria

- The user (and I) have a shared understanding of when the cost is
  justified.
- Measurable reduction in conversations spent on tasks that don't need
  an LLM.

## Honest assessment

High value but requires discipline from both sides. The framework itself
is cheap to create. The hard part is actually following it — especially
when the LLM is convenient even for tasks that don't need it. This plan
is more about establishing a norm than building a tool.