
This is the manual you need when Claude starts feeling slow or forgetting things mid-session. It covers the basics (run /compact at 70% context usage, delegate heavy output to subagents) and the less obvious stuff (each MCP adds overhead to every request, keep it under 10 total). The token efficiency section is especially good: no sycophantic openers, code first, and switching response styles based on whether you're coding versus running agent pipelines. The 40-60% output savings from cutting fluff alone makes it worth reading. If you've ever hit the "dumb zone" at 90% context or burned through budget faster than expected, the quick diagnosis table and task profiles will save you real money.
$ npx -y skills add rohitg00/pro-workflow --skill context-optimizer --agent claude-codeInstalls into .claude/skills of the current project.
Manage your context window and token budget effectively.
/context to check current usage| Action | Saves | When |
|---|---|---|
/compact | 30-50% context | At task boundaries |
| Disable unused MCPs | ~5% per MCP | When switching domains |
| Use subagents for exploration | Keeps main context clean | Heavy search/read tasks |
Fresh session via /resume | 100% reset | When starting unrelated work |
Set proactive auto-compaction:
{
"env": {
"CLAUDE_AUTOCOMPACT_PCT_OVERRIDE": "50"
}
}
Keep <10 MCPs enabled, <80 tools total. Each MCP adds overhead to every request.
/mcp # List active servers
# Disable what you're not using
Sponsors