Reading your first audit

Running /token-optimizer once after install gives you a complete picture of where your context goes and a guided path to recover it. The audit runs in phases: it scans your setup with parallel agents, synthesizes the findings, presents them on a dashboard, and applies the fixes you approve. Nothing changes without your sign-off, and every change is backed up first.

Typical recovery is 5 to 15% of your context budget, and up to 25% or more once autocompact management is in place.

What the audit does, phase by phase

1. Scan. Six auditors run in parallel, one per source of structural waste: CLAUDE.md, MEMORY.md, skills, MCP servers, slash commands, and settings or hooks. Each measures size, duplication, and dead weight in tokens.

2. Synthesize. A synthesis pass reads every auditor’s findings and produces one ranked optimization plan: what to fix, how many tokens each fix recovers, and the side effects to watch for.

3. Present. The plan renders on the dashboard so you can see the per-component breakdown and the projected savings before deciding anything. The audit waits here for you.

4. Implement. For each fix you approve, Token Optimizer shows a before-and-after diff, backs up the original, and applies the change. Fixes cover trimming CLAUDE.md, archiving stale MEMORY.md entries, archiving unused skills, disabling dead MCP servers, adding file-exclusion rules, tuning hooks and caching, and installing compact instructions.

5. Verify. A final pass re-measures everything and reports the actual before-and-after savings, plus behavioral next steps for the habits a single audit cannot fix.

How to act on the findings

Read the findings as a budget, not a dollar figure. Each item tells you how many tokens it recovers and what share of your context window that is. Prioritize the structural items at the top of the plan: they recover the most and they compound, because a leaner prefix lowers your cache-read bill on every turn that follows.

Approve fixes one at a time when you are unsure. The audit presents a diff for each change and waits for you. You can approve the safe, high-recovery items (duplicate configs, orphaned memory entries, unused skill frontmatter) and defer anything that touches a skill, MCP server, or deny rule another tool depends on.

Note

The audit never deletes files. Skills and configs are archived outside the skills directory so you can restore them. It also checks dependencies before archiving, because a skill, MCP server, or deny rule can be load-bearing for something else.

After the audit

The structural fixes are one-time. The rest runs on its own: active compression cuts runtime waste as it enters your window, Smart Compaction protects your work across auto-compact, and the quality score tracks degradation turn by turn. The dashboard regenerates after every session so you can watch the savings accumulate.

For the behavioral patterns the audit surfaces (compacting too late, model misrouting, sessions creeping longer), run Token Coach, which reads 30 days of your session history.

Next steps

• The dashboard: how to read every panel.
• Active compression overview: what runs automatically after the audit.
• Managing skills and MCP: toggle components on and off after the audit.