| name | context-budget |
| description | Audit token costs of the .claude/ setup — skills, agents, hooks, and CLAUDE.md. Shows what's always loaded vs on-demand, flags bloat, and suggests cuts. |
Context Budget Audit
Run the audit script and report findings:
Execute: bash ../context-budget.sh
Then analyze the output:
-
Classify each item
- Always loaded: hook outputs (SessionStart), CLAUDE.md
- On invocation: skill SKILL.md files (when
/skill called)
- On Agent spawn: full agent .md body
-
Flag if over budget
- CLAUDE.md > 500 tokens → summarize or split
- Any single skill > 200 tokens → trim
- Agent description line > 50 tokens → shorten
- SessionStart output > 100 tokens → slim hook script
-
Report format
ALWAYS LOADED
CLAUDE.md ~XXX tokens
SessionStart.sh ~XX tokens (estimated output)
ON INVOCATION (skills)
/build ~XX tokens
/format ~XX tokens
...
ON AGENT SPAWN
graphql-explorer ~XXX tokens
TOTAL LOADED AT REST: ~XXX tokens
RECOMMENDATIONS: ...