context-engineer

# Context Engineer

Audit context consumption and recommend compaction strategies.

## Trigger

Use when sessions feel slow, compaction fires too often, token budget seems wasted, or before adding new skills/agents/MCP servers.

## Workflow

1. Measure CLAUDE.md total size (root + all descendants)
2. Count installed skills and estimate description overhead
3. Count MCP servers and their tool count
4. Count agents and their preloaded skills
5. Identify redundancy across memory files
6. Score context health (0-100)
7. Recommend specific actions

## Audit Checklist

### CLAUDE.md Size
- Root CLAUDE.md: ideal < 60 lines, maximum < 150 lines
- Total across all CLAUDE.md files: flag if > 500 lines
- Check for stale entries, duplicate info, verbose examples

### Skill Overhead
- Each skill description consumes ~50-200 tokens in the slash menu
- Flag skills that haven't been invoked in 30+ days
- Identify skills that overlap in purpose

### MCP Servers
- Each server adds tool descriptions to context
- Count total tools across all servers
- Flag servers with 20+ tools (consider tool filtering)

### Agent Preloaded Skills
- Skills in agent frontmatter load fully into agent context
- Flag agents with 3+ preloaded skills
- Suggest moving rarely-needed knowledge to on-demand skills

## Output

```
CONTEXT AUDIT
Date: [date]
Health: [score]/100

CLAUDE.md: [X] lines across [N] files
  Root: [X] lines ([ok/bloated])
  Largest: [path] ([X] lines)

Skills: [N] installed, ~[X] tokens in descriptions
  Unused (30d+): [list]
  Overlapping: [pairs]

MCP Servers: [N] servers, [X] total tools
  Heaviest: [name] ([X] tools)

Agents: [N] agents, [X] preloaded skills total

RECOMMENDATIONS:
1. [Highest impact action]
2. [Second action]
3. [Third action]

MOVE TO SUB-AGENTS:
- [Knowledge that doesn't need to be in main context]

KEEP IN MAIN CONTEXT:
- [Knowledge needed every session]
```

## Rules

- Never modify files. Read-only analysis.
- Prioritize recommendations by token savings.
- Always distinguish between "move to sub-agent" and "delete entirely."
- Flag anything over 200 lines in a single CLAUDE.md as a split candidate.