From agent-caveman
Use when spawning subagents, delegating research, or calling WebFetch. Picks the right specialized subagent over general-purpose, and forces tight extraction prompts on web fetches. Triggers on "multi-agent workflow", "delegate", "spawn subagent", "research this", "fetch", "summarize this URL", "look up".
How this skill is triggered — by the user, by Claude, or both
Slash command
/agent-caveman:grunt-orchestratorThe summary Claude sees in its skill listing — used to decide when to auto-load this skill
Two jobs, both source-side token control:
Two jobs, both source-side token control:
Pick the narrowest role. general-purpose is the fallback, not the default.
| Task pattern | Use |
|---|---|
| "look up docs", "find prior art", web search | grunt-researcher |
| "where is X in the code", "how does X work" | grunt-explorer |
| "edit file Y to do Z" (with spec ready) | grunt-coder |
| Task spans roles or needs all tools | general-purpose |
| Task touches plugins/hooks/skills/API details | claude-code-guide |
Each grunt-* agent ships a strict output contract (RESULT / FINDINGS / FILES / NEXT) — you'll get structured facts, not prose. That's the point.
When briefing any subagent:
max 200 words)Data from this project: WebFetch averages ~7,200 output tokens per call when prompts are vague. Tight prompts cut that by an order of magnitude — the underlying small model generates only what you asked for.
Rule: ask for the extract, not a summary of the page.
| Bad prompt | Better prompt |
|---|---|
| "What does this page say about plugins?" | "Return only the minimal plugin.json required fields as a JSON code block. No prose." |
| "Tell me about the MCP protocol." | "List the exact names of the four MCP primitives (tools/resources/prompts/sampling). Name only, one per line." |
| "Summarize Anthropic's hook documentation." | "Return the exact JSON shape a PostToolUse hook emits to mutate tool output. Code block only." |
| "What's the install command?" | "Return only the shell command to install this as a Claude Code plugin. One line." |
Patterns that keep WebFetch responses small:
Before calling WebFetch, ask yourself: what would the ideal response look like? Then demand exactly that.
general-purpose when grunt-explorer fits — loads every tool's schema into the subagent's contextGuides creation, editing, and verification of skills for AI coding agents using test-driven development with subagent scenarios. Use when authoring or debugging skills.
npx claudepluginhub carlet0n/agent_caveman --plugin agent-caveman