Dynamic Hive Skill — AI skill for Claude Code

Hive — General-Purpose Multi-Agent Orchestration

Hive dynamically designs and executes a swarm of specialist AI agents for **any task type** — research, writing, strategy, analysis, data, creative work, or hybrid projects. It selects the right model and execution harness per agent based on benchmarks and cost, never habit or stereotype.

---

Core Principle

Assign the **minimum-sufficient model** that can fulfill a role's cognitive demands. Justify every pick from benchmarks and cost. Step up exactly one tier on demonstrated failure.

Multi-agent quality gains come from matching each agent's **role, model, and harness** to the specific demands of its task — not from using more agents or more powerful models uniformly.

---

When to Use Hive

**Use it when:**

• The task has 3+ distinct components that must interoperate
• The task is too large for one agent's context window
• You need parallel workstreams with a unified final deliverable
• User says: "build this", "research this thoroughly", "create a comprehensive X", "multi-agent", "swarm"

**Do not use it for:**

• Single-agent tasks with no interdependencies
• Tasks completable in one step
• Quick fixes or one-file changes

---

The Seven Phases

Phase 1: Task Analysis

Read `resources/task-taxonomy.md`.

Classify the task into one of five families: **Knowledge, Construct, Transform, Decide, or Hybrid**. Calculate complexity (1–25 scale) and a coupling score (0.0–1.0) measuring how tightly components depend on each other.

Output format:

HIVE ANALYSIS
=================
Task Family: [Knowledge|Construct|Transform|Decide|Hybrid]
Objective: [one sentence]
Components: [list]
Complexity: [1-25]
Coupling Score: [0.00]
Recommended Mode: [Parallel|Hybrid|Sequential]

Phase 2: Coordination Mode Selection

Read `resources/coordination-modes.md`.

Coupling Score	Mode	Use When
< 0.4	Contract-First Parallel	Independent components — 4–8x speedup
0.4–0.7	Hybrid Wave