| name | grok-audit-context |
| description | Use when the user asks what's loaded in the current Grok Build session, why context feels heavy, how many tokens subagents are using, or how to reduce overhead. Covers Grok-specific concepts: injected system prompt size, available skills (from ~/.grok/skills + project .grok/skills + bundled), active MCP servers (grok_com_github etc.), subagent types (implement, review, explore, plan, best-of-n), and the workspace path. Recommends trims. Complements (does not duplicate) the Claude `audit-context` skill. |
Grok Build Context Audit
Inspect the current Grok Build / Grok Heavy session for context bloat and recommend targeted reductions.
Grok Build has a different loading model than Claude Code:
- Large system prompt injected on every turn (includes full tool schemas + skill metadata).
- Subagents (implement, review, explore, plan, design, best-of-n) each carry their own persona + instructions.
- Skills are discovered from multiple locations and listed in the system reminder.
- MCP servers appear as tools when connected.
What to Capture
1. Workspace & Identity
echo "Workspace Path (from runtime): $PWD"
echo "Effective develop root: /develop or /Users/sunrunner/Develop"
echo "Dotfiles: ~/dotfiles (source of truth for GROK.md + skills)"
2. Skills visible to this session
Look at the system-reminder block that lists "Available skills". Count them and group by source:
- Bundled skills (
~/.grok/bundled/skills/) — implement, review, best-of-n, design, pr-babysit, etc. (heavy but very capable)
- User skills (
~/.grok/skills/)
- Project skills (
<cwd>/.grok/skills/ or walking up to ~/Develop/.grok/skills/)
- Any
.claude/skills/ that the environment also surfaced (hybrid visibility)
Report rough token cost estimate: number of skills × 60–120 tokens each at session start.
3. Active MCP servers
Note which ones are global vs project-scoped.
4. Subagent overhead
Grok Build's power comes from subagents. Each one carries:
- Full persona prompt (planner, coder, reviewer, researcher, preservation guard, etc.)
- Access to the same tool set + skills
When the user has run implement, review, or multiple Explore agents, context can grow quickly because each subagent maintains its own history.
5. Current working directory scope
Confirm we are under /develop (or /Users/sunrunner/Develop).
If the user is working in a single cerid-* repo, having the full set of Develop-scoped skills + all cerid MCPs loaded may be unnecessary.
Recommendations (Grok-flavored)
- Disable unused bundled skills (if the user rarely uses
design, pr-babysit, best-of-n, etc.). These are the heaviest.
- Prefer project-scoped skills over global ones for cerid-specific or HA-specific workflows.
- Use
implement skill only when the task genuinely benefits from the full reviewer loop — it is powerful but expensive.
- For pure exploration, use a single
explore subagent or manual parallel tool calls instead of forking many Explore agents.
- MCP scoping: If
cerid-kb or HA tools are appearing in unrelated cerid repos, check for a leaky parent .grok/config.toml or .mcp.json.
Output Format
Produce a clear markdown report with:
- Current workspace
- Breakdown of loaded skills by source + rough token estimate
- Active MCP servers
- Subagent usage pattern observed in this session
- Concrete, ordered trim recommendations (e.g. "move X skill from global to project", "disable bundled skill Y in ~/.grok/config.toml", "add .grokignore equivalent if one exists")
Do not automatically edit config files — present the diffs and let the user approve.
Out of Scope
- Does not analyze the user's application code performance.
- Does not touch Claude Code session state (use the Claude
audit-context skill for that).
