Compactor Skill

Token Saver

Reduce Claude Code token usage without altering functionality. This skill enforces efficient patterns, suggests model switching, and monitors context growth.

When to activate

• ALWAYS active — these rules apply to every conversation
• Invokable via /token-saver for a mid-conversation efficiency check

Core Rules (ALWAYS FOLLOW)

1. Targeted Reads — NEVER read full files blindly

• ALWAYS use offset + limit when you know which section you need
• Use Grep FIRST to find the relevant line numbers, THEN read only those lines
• For files >200 lines: read the first 30 lines to understand structure, then target specific sections
• BAD: Read file.py (reads 2000 lines, burns tokens)
• GOOD: Grep "def process_data" file.py → Read file.py offset=145 limit=30

2. Model-Appropriate Work

Suggest model switches to the user when appropriate:

• Haiku (/model haiku): grep, file lookups, formatting, simple Q&A, running commands, generating boilerplate, SQL queries, status checks
• Sonnet (/model sonnet): coding tasks, bug fixes, code review, moderate analysis, writing tests
• Opus (/model opus): complex architecture, multi-file refactors, deep debugging, planning, nuanced analysis

When the task is simple (lookup, grep, format), say: "This is a quick lookup — consider /model haiku to save tokens."

3. Concise Responses

• Lead with the answer, not the reasoning
• Max 3-5 sentences for simple questions
• No preamble ("Let me...", "I'll now...", "Sure, I can...")
• No trailing summaries ("In summary...", "To recap...")
• Code blocks: only show the changed portion, not the entire file
• Tables: only when genuinely clearer than prose

4. Smart Tool Usage

• Use Glob not find (dedicated tool = fewer tokens in output)
• Use Grep not cat file | grep (dedicated tool)
• Use Edit not Write for changes (sends diff, not full file)
• Batch independent tool calls in a single message (parallel execution)
• NEVER spawn an Agent for something that takes 1-2 tool calls
• Agents are expensive — each gets its own context. Only use for genuinely parallel, complex, multi-step work

5. Context Management

• After 15+ back-and-forth messages, suggest /compact to the user
• After completing a major task, suggest /compact before starting the next one
• When conversation has accumulated many large tool outputs, suggest /compact
• Phrase: "Context is getting large — /compact would save tokens for the rest of this session."

6. Avoid Waste

• Don't re-read files you already read in this conversation (use your memory of the content)
• Don't re-run commands that already succeeded unless state changed
• Don't explore code you don't need for the current task
• Don't add unsolicited improvements, refactors, or documentation
• If a task is done, stop. Don't offer "would you also like me to..."

When invoked via /token-saver

Run a mid-conversation efficiency check and report:

## Token Efficiency Check

### Current Session
- Messages exchanged: ~{count}
- Large file reads: {list any full-file reads that could have been targeted}
- Agent spawns: {count} — were they all necessary?
- Model: {current} — appropriate for current work?

### Recommendations
- [ ] Run /compact (context is large)
- [ ] Switch to /model haiku (current task is simple)
- [ ] Next reads: use offset+limit, not full file

CLAUDE.md Integration

When this skill is first loaded, check if ~/.claude/CLAUDE.md or the project CLAUDE.md contains token efficiency rules. If not, suggest adding:

# Efficiency
- Prefer targeted file reads (offset+limit) over full reads
- Use Grep before Read to find line numbers
- Batch independent tool calls in parallel
- Suggest /compact after major task completion
- Keep responses concise — answer first, explain only if asked

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: chethanbhatbs
• Source: chethanbhatbs/compactor-skill
• License: MIT

Install and usage instructions live in the source repository linked above.