dragonkid/claude/skills/llm-agent-token-audit

LLM Agent System Token Audit

Systematic audit for reducing fixed per-call token overhead in LLM agent frameworks that inject workspace files into system prompts.

When to Use

• Per-call token usage feels high or costs are growing
• Agent framework injects workspace files (Claude Code, custom setups)
• Setting up new agent and want lean defaults
• After adding skills/tools, want to check overhead

Claude Code users: For auditing Claude Code's own context (skills, agents, rules, MCP servers), use /context-budget instead — it scans Claude Code components with specific token thresholds. This skill is for auditing custom agent frameworks and general LLM setups.

Audit Steps

1. Quantify

Find the framework's prompt report (/context detail, sessions.json, equivalent). Record exact char counts per injected component.

2. Classify Waste

| Type | Signal | Fix | |------|--------|-----| | Dead files | One-time files still present | Delete | | Template bloat | Unfilled placeholders/examples | Clear or fill with real content | | Redundancy | Same info in multiple files | Deduplicate | | Unused features | Skills/tools loaded but never used | Disable via config | | Model mismatch | Expensive model on simple tasks | Reassign per-task |

3. Prioritize

• P0: Delete/clear (zero risk, immediate savings)
• P1: Config changes (disable skills, adjust settings)
• P2: Content rewrite, model tiering (more effort)

For file-level content optimization, use optimize-llm-config-docs skill.

4. Execute Safely

• Version-control before modifying
• Critical instruction files: edit incrementally with review
• Non-critical (templates, dead files): batch delete/clear
• Restart agent after config changes
• Verify with framework's context reporting tool

Common Mistakes

• Rewriting critical instruction files in one shot without review
• Optimizing by gut feel instead of measuring first
• Disabling skills that are rare but critical
• Forgetting to restart after config changes