Vibe Aigc — Development skill for Claude Code

vibe-aigc

Implementation of the Vibe AIGC paper ([arXiv:2602.04575](https://arxiv.org/abs/2602.04575)). Decomposes high-level creative intent ("Vibes") into executable multi-step content generation workflows via agentic orchestration.

What it does

You describe what you want (a "Vibe" -- style, constraints, description). A MetaPlanner breaks it into a DAG of tasks, assigns tools, executes nodes (with parallelism where possible), and returns composed output. If a node fails, it replans.

from vibe_aigc import Vibe, MetaPlanner

vibe = Vibe(
    description="Write a blog post about AI agents",
    style="informative, engaging",
    constraints=["under 1000 words"]
)

planner = MetaPlanner()
result = await planner.execute(vibe)

Architecture

Based on the paper's 3-component design:

Component	Role	Module
MetaPlanner	Decomposes Vibes into workflow DAGs	vibe_aigc.planner
KnowledgeBase	Domain expertise (film, writing, design, music)	vibe_aigc.knowledge
ToolRegistry	Atomic content generation tools	vibe_aigc.tools

Additional modules handle ComfyUI workflow composition, video/audio pipelines, VLM feedback loops, and multi-GPU execution.

LLM backends

Provider	Config	Key required
OpenAI	OPENAI_API_KEY env var	Yes
Anthropic	ANTHROPIC_API_KEY env var	Yes
Ollama	LLMConfig.for_ollama(host=..., model=...)	No

Auto-detection checks env vars in order, falls back to local Ollama.

Install

pip install vibe-aigc

Requires Python 3.12+.

CLI

vibe-aigc plan "Create a blog post about AI" --style informative --format ascii
vibe-aigc execute "Design a landing page" --visualize --checkpoint
vibe-aigc checkpoints --list

Features

• Parallel execution of independent workflow nodes
• Checkpoint/resume for long-running workflows
• Adaptive replanning on node failure
• ASCII and Mermaid workf