grunt-orchestrator

Grunt Orchestrator

Two jobs, both source-side token control:

1. Route delegation to the tightest subagent that fits
2. Write WebFetch / research prompts that ask for the answer, not the page

1. Subagent routing

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:

• State the goal in one sentence
• List what's already ruled out
• Hand over exact paths/URLs/commands when known
• Cap response length explicitly (max 200 words)
• State whether code changes are expected or read-only

2. WebFetch prompts

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:

• "Return only..." / "One line." / "No prose." / "Code block only."
• Ask for a specific field, schema, command, or URL — not an explanation
• If you need prose, cap words: "Max 40 words, no preamble."
• Name the output format: JSON, table, bullet list of ≤5 items

Before calling WebFetch, ask yourself: what would the ideal response look like? Then demand exactly that.

Anti-patterns

• Spawning general-purpose when grunt-explorer fits — loads every tool's schema into the subagent's context
• Asking a subagent to "explore and tell me what you find" — returns narrative, not facts
• WebFetch prompts phrased as questions ("What is X?") — the model answers conversationally, long
• Forgetting to cap response length when delegating — subagents default to thorough prose

When to ignore this skill

• Task is trivial, single tool call, done inline — no delegation needed
• User explicitly asked for full prose / teaching tone — compression disserves them
• Security warnings, destructive ops, confirmations — clarity beats brevity