From noxfen-essentials
Multi-agent task orchestration skill. Use when the user asks to "orchestrate", "spawn agents", "parallelize this", "plan with subagents", "break this into agents", or when a task is clearly complex enough to benefit from parallel subagents (large refactors, multi-module features, cross-cutting analysis). Also triggers on "/orchestrate" or "/swarm". Analyzes complexity, token budget, and routes work to the right models to minimize cost.
How this skill is triggered — by the user, by Claude, or both
Slash command
/noxfen-essentials:orchestratorThe summary Claude sees in its skill listing — used to decide when to auto-load this skill
Decompose complex tasks into parallel subagent workstreams. Route each workstream to the cheapest
Decompose complex tasks into parallel subagent workstreams. Route each workstream to the cheapest model that can handle it. Synthesize results in the main context.
Before spawning anything, classify the task:
| Level | Criteria | Action |
|---|---|---|
| Simple | Single file, clear scope, <200 lines affected | Do it directly, no agents |
| Medium | 2–5 files, clear interfaces, isolated changes | 1–2 agents max |
| Complex | Cross-cutting, >5 files, multiple concerns | 3–5 specialized agents |
| Massive | Full feature, architectural change, >10 modules | Phase into waves; first wave scopes, second executes |
If Simple → stop here and do the task directly.
Estimate token cost before spawning:
Current model costs (relative, as of 2026):
Rule: 80% of token work should land on Haiku/Sonnet. Opus only for what truly needs it.
Decompose the task into independent workstreams. Each must be:
Explorer (Haiku): Read-only codebase analysis. Find files, grep symbols, map dependencies. Prompt pattern: "Search X in codebase. Report: file paths, line numbers, relevant snippets. No edits."
Implementer (Sonnet): Write/edit code for a single bounded module or concern. Prompt pattern: "Implement X in files [list]. Context: [what explorer found]. Return: diff summary."
Reviewer (Sonnet): Check a completed implementation for correctness, edge cases, consistency. Prompt pattern: "Review [what implementer did]. Check: [specific concerns]. Report issues or OK."
Architect (Opus): Only for genuinely ambiguous design decisions or cross-cutting concerns. Use sparingly — one per session max unless budget is large.
Spawn all independent agents in a single message (parallel). For dependent workstreams, wait for upstream agents before spawning downstream ones.
Before spawning, state the plan to the user:
Orchestrating N agents:
[Agent 1] Explorer → [scope] (Haiku)
[Agent 2] Implementer → [module A] (Sonnet)
[Agent 3] Implementer → [module B] (Sonnet)
Estimated tokens: ~Xk. Proceed? [y/n]
Wait for user confirmation before spawning if total estimated cost exceeds 100K tokens.
After all agents complete:
Main context is the scarcest resource. Every byte the orchestrator adds persists for the entire session.
When prompting each subagent, always end with:
Report in max 200 words: what changed, what issues found, what to do next.
No reasoning chain. No code blocks unless the exact code is the answer.
When collecting results:
If context estimate (from statusline ctx %~) approaches 70%:
/compact before the next major taskAnti-patterns:
npx claudepluginhub noxfen/claude_skill_everywhere --plugin noxfen-essentialsCreates, edits, and optimizes skills for Claude Code, including drafting, evaluating with test prompts, iterating on performance, and improving skill descriptions for better triggering accuracy.