Support Pi and Other Agent Runtimes

[shenxiuqiang]

Background

The current ArcBlock/agent-skills repository provides a useful collection of ArcBlock-related Agent Skills, including arc, blocklet, arcblock-context, content-creation, and other capabilities.

At the moment, the repository is primarily distributed and integrated as a Claude Code plugin. This works well for Claude Code, but the underlying SKILL.md format is increasingly becoming a common way to share reusable agent capabilities across different AI coding agents.

For example, Pi (pi.dev) already supports the Agent Skills standard and can discover skills from directories containing SKILL.md. Other agent runtimes are also adopting similar skill-based architectures.

This creates an opportunity to make ArcBlock/agent-skills more portable and useful across the growing agent ecosystem.

Proposal

I suggest restructuring or extending the repository so that the actual Agent Skills are kept agent-runtime independent, while providing lightweight integration layers for different agents.

For example:

agent-skills/
├── skills/
│   ├── arc/
│   │   └── SKILL.md
│   ├── blocklet/
│   │   └── SKILL.md
│   ├── arcblock-context/
│   │   └── SKILL.md
│   └── ...
│
├── .claude-plugin/
│   └── ...
│
├── package.json
└── README.md

The core SKILL.md files would contain the actual knowledge and instructions, while Claude Code, Pi, Codex, and other agent runtimes could provide their own integration/discovery mechanisms.

Pi support

Pi already supports Agent Skills through directories such as:

~/.pi/agent/skills/
~/.agents/skills/
.pi/skills/
.agents/skills/

It would therefore be useful if the ArcBlock skills could be consumed by Pi without requiring a Claude Code-specific conversion.

For example, ideally users could install the repository with something similar to:

pi install git:github.com/ArcBlock/agent-skills

and then use:

/skill:arc
/skill:blocklet

or discover the available ArcBlock skills through Pi's normal skill mechanism.

If direct package installation is not appropriate, simply documenting how to point Pi at the repository's SKILL.md directories would already provide significant value.

Broader Agent Compatibility

I would also suggest treating Agent Skills as the portable core, rather than making the skills themselves dependent on a particular agent runtime.

Potential integrations could include:

• Claude Code
• Pi
• Codex
• Cursor
• Gemini CLI
• Other agents supporting the Agent Skills standard

The architecture could therefore be:

                  ArcBlock Agent Skills
                          │
             ┌────────────┼────────────┐
             │            │            │
        Claude Code       Pi          Codex
             │            │            │
          Plugin       Skills        Skills
             │            │            │
             └────────────┼────────────┘
                          │
                   ArcBlock knowledge

This would allow ArcBlock to maintain the skill content in one place while making it available to users regardless of which AI agent they prefer.

Why This Matters

AI coding agents are becoming increasingly fragmented. Developers may use Claude Code, Codex, Pi, Cursor, or other agents depending on the task and workflow.

If ArcBlock skills are tightly coupled to one agent, developers using other agents need to manually copy, convert, or recreate the skills.

By keeping the core skills runtime-independent, ArcBlock could instead provide a reusable ArcBlock Agent Skills ecosystem.

This would also make the repository easier for the community to contribute to: developers could create or improve a skill once, rather than maintaining separate implementations for different agents.

Suggested Next Step

I suggest considering the following:

1. Identify the existing SKILL.md files and separate them from Claude-specific plugin metadata.
2. Establish a runtime-independent skills/ directory as the canonical source of ArcBlock skills.
3. Keep the existing Claude Code plugin integration as one distribution layer.
4. Add Pi support through its native skill/package mechanism.
5. Document compatibility with other Agent Skills-compatible runtimes.
6. Consider adding CI validation to ensure every skill follows the expected Agent Skills structure.

This would preserve the current Claude Code experience while making the repository a more general cross-agent skill distribution platform for ArcBlock.

I believe this direction could also make agent-skills more future-proof as the agent ecosystem continues to evolve.