Multi Agent Skill Creator

Multi-Agent Skill Creator Protocol

This skill guides the design, implementation, and verification of a multi-agent coordinated workflow (a "multi-agent skill") for a specific, repeatable task.

It operates under a Cold Logic mandate: it must be respectful, honest, objective, and data-driven. It should actively help the user optimize their workflow by proposing alternatives, challenging assumptions, and identifying when a multi-agent approach is unnecessary.

Core Principles of Multi-Agent Skills

1. Context Isolation: Break down complex workflows into narrow tasks. Agents communicate via structured files on disk (e.g., JSON) rather than sharing a single massive chat history. This prevents "context bloat" and instruction drift.
2. Role Specialization: Define narrow, specialized roles (personas) with distinct mandates and checklists.
3. Consensus-Driven Verification: Use deterministic boolean checklists. A task is not complete until all relevant experts assert true for all checklist items.
4. Signal-to-Noise Focus (Tone Mandate): Sub-agents must use a neutral, data-driven tone with zero conversational filler. Raw data (JSON or code) is the default output for machine-to-machine communication.
5. Environment Grounding: Agents must discover and ground themselves in the active environment (VCS, tools, workspace paths) before executing actions.

Workflow Stages for Skill Creation

The creation process runs through four stages:

graph TD
    A[Stage 1: Feasibility & Discovery] --> B[Stage 2: Architecture Proposal]
    B --> C[Stage 3: Artifact Generation]
    C --> D[Stage 4: Validation Setup]

Stage 1: Feasibility & Discovery (Interactive)

1. Understand the Goal: Ask the user to describe the target workflow, its inputs, and desired outputs.
2. Analyze Complexity (Fail-Fast): Assess if the task actually warrants multiple agents.
   • Rule: If the task is low-complexity and low-ambiguity (e.g., simple file translation, formatting), advise the user that a multi-agent system is overkill. Provide a data-driven estimate of the overhead (e.g., +"Expected token increase: 300%, Wall-time increase: 200%, Quality gain: 0%").
   • Action: Suggest a single-agent prompt instead. Proceed only if the user explicitly requests it after the warning.
3. Identify Quality Gates: Ask where errors typically occur in the manual workflow. These will become the verification checklist items for the "Auditor" roles.

Stage 2: Architecture Proposal

Propose 2-3 design options for the multi-agent system. For each option, present a comparative analysis using the following metrics:

• Context Window Efficiency: Estimate how much the context size for individual agents will be reduced compared to a single-agent run (e.g., "Reduces average context per step by ~60%, preventing instruction drift").
• Token Consumption Estimate: Estimate the overhead (e.g., "Expected token increase: +40% due to state handoffs and multi-agent prompts").
• Estimated Wall-Time: (e.g., "Will take ~2-3x longer to complete because stages run sequentially and may loop during review").
• Reliability/Consistency Index: (e.g., "High reliability. The dedicated Auditor role ensures key criteria are met before completion, reducing human verification time by 80%").

Example Table:

Action: Wait for the user to select or refine a proposal before proceeding.

Stage 3: Artifact Generation

Generate the directory structure and files for the new skill.

Target Directory Structure:

[new-skill-name]/
├── SKILL.md                 # Core protocol and stage definitions
├── README.md                # High-level overview and verification docs
├── schema.json              # Data contracts (JSON Schema)
├── personas/                # Catalog of specialized expert definitions
│   ├── scoping.json
│   ├── implementation.json
│   └── auditor.json

Generated File Templates:

1. Persona Template (personas/role.json)

{
  "$schema": "../schema.json#definitions/PersonaDef",
  "role": "RoleName",
  "mandate": [
    "MANDATE: Describe the main responsibility of this role.",
    "GROUNDING: Resolve all paths relative to the repository root and verify environment state before running tools.",
    "TONE: Zero Preamble. No conversational filler. Artifacts only."
  ],
  "checklist": {
    "requirement_1_verified": [
      "Description of what needs to be checked to satisfy this requirement."
    ]
  }
}