AI-Rig: A Professional AI Programming Assistant Configuration System for Python/ML Open-Source Projects

This post introduces AI-Rig, a specialized configuration system designed for Python and machine learning open-source projects. It features 12 professional agents, over 20 workflow skills, and a collaborative integration scheme for Claude Code and Codex CLI. Unlike general AI assistants, AI-Rig adopts a "professional agents for professional tasks" philosophy, building a team of domain-expert agents to handle complex development workflows, aiming to enhance efficiency and code quality for open-source maintainers and ML researchers.

General AI assistants often lack depth in professional domain tasks. AI-Rig was created to address this gap—it's a carefully designed system optimized for Python/ML open-source development. Its core philosophy: "Professional problems require professional agents." Instead of a single general assistant, it constructs a team of domain experts, each focusing on their area of expertise to collaborate on complex workflows.

AI-Rig uses a modular plugin architecture with four core layers:

1. Foundry (Base Layer): Foundation with 8 basic agents (doc-scribe, linting-expert, perf-optimizer, etc.) and core skills.
2. OSS (Open-Source Workflow Layer): For maintainers, offering governance workflows like issue classification, PR review, and CI management.
3. Develop (Development Layer): For daily tasks, supporting TDD-based workflows (feature dev, bug fixes, refactoring, debugging, planning).
4. Research (Research Layer): For ML research, providing experiment-driven workflows (paper analysis, dataset management).

AI-Rig's core innovation is its "Skill" system—skills are complete workflows defining multi-agent collaboration (participants, execution order, info transfer). Examples:

• Code Review Skill (/oss:review): 7-agent layered process (Tier0: script diff stats, Tier1: Codex pre-review, Tier2: Claude deep analysis).
• Feature Dev Skill (/develop:feature): Strict TDD flow (codebase analysis → demo test → red-green-refactor → docs update → quality review). Skills support chained calls (e.g., problem analysis → fix → validation; performance investigation → optimization → refactoring).

AI-Rig integrates Claude Code and Codex CLI with clear division:

• Claude Code: Long-cycle reasoning, workflow orchestration, decision-making.
• Codex CLI: Focused mechanical code tasks (direct shell operations). Three-tier review pipeline: Tier0 (script, sec-level stats), Tier1 (Codex, ~60s diff review), Tier2 (Claude agents, minute-level deep analysis). Claude delegates mechanical tasks to Codex (conflict resolution, applying review comments, batch refactoring). A real case showed combined perspectives are more comprehensive (Claude's style unification + Codex's blind spot detection).

AI-Rig delivers value for key scenarios:

1. Open-Source Maintainers: Streamlines tasks like issue classification (/oss:analyse), PR review (/oss:review), bug fixes (/develop:fix), and version releases (/oss:release).
2. ML Researchers: Supports experiment workflows (literature research /research:topic, experiment planning /research:plan, automated runs /research:run).

Future plans: 1) Agent self-improvement (via /distill//calibrate), 2) Memory management (clean outdated entries), 3) MCP server integration (OpenSpace for skill evolution, colab-mcp for GPU workloads), 4) Token optimization (RTK tool for shell output compression).

Conclusion: AI-Rig is a well-designed agent ecosystem, not just prompts. It elevates AI assistants from chat tools to collaborative partners, maximizing human-AI efficiency for Python/ML open-source projects.