DEV Community

Fenju Fu
Fenju Fu

Posted on

Agent Skill Packs Are Booming — But They're All Coding Skills. What About Multimodal?

#aiagents #multimodal #ocr #skills

Today's GitHub Trending tells a clear story: agent skill packs are hot. Three repos on the list — addyosmani/agent-skills (84K+ stars), google/skills (16K+ stars), and mattpocock/skills (209K+ stars) — all package skills that AI coding agents can use.

The skills they cover? Code review, refactoring, test generation, clean architecture patterns, dependency management. All coding. All engineering.

This makes sense. The first wave of AI agents was coding agents — Cursor, Copilot, Windsurf — so the first wave of skill packs naturally serves that use case.

But here's the gap: what about agents that work in offices, not IDEs?

The Multimodal Skill Gap

Imagine an office agent that receives: "Process these invoices, translate the contract from Chinese to English, and proofread the report before submission."

What skills does it need?

  • OCR: Extract text from tilted, photographed, or scanned invoices — not clean PDFs, real-world paper.
  • Translation: Handle contracts with nested tables, mixed languages, and domain-specific terminology.
  • Proofreading: Catch typos, grammar issues, and formatting inconsistencies in formal documents.
  • Speech: Transcribe meeting recordings, convert voice notes to structured text.

None of these are coding skills. They're multimodal capabilities — and there's no trending skill pack for them.

iFly-Skills multimodal capabilities overview

Why Multimodal Skills Are Harder to Package

Coding skills are text-in, text-out. A refactoring skill is a prompt template with some context. Relatively easy to package, version, and share.

Multimodal skills are different:

  • OCR needs to handle image quality, tilt, lighting, and layout complexity.
  • Translation needs domain models (legal, medical, technical) — not just general-purpose NMT.
  • Proofreading needs to understand document structure, not just string matching.

This is why most skill pack repos skip multimodal: it's harder to make a generic prompt template for "OCR this tilted invoice" than for "refactor this function."

The Official + Vertical Approach

Two patterns from today's Trending are worth noting:

  1. "Production-grade" trust anchor (addyosmani): Labeling skills as "production-grade" signals "not a toy." Developers trust it.
  2. Official backing (google/skills): Google's official label builds instant trust. Developers prefer "official" over "community-sourced" for critical capabilities.

Both patterns apply to multimodal skills — arguably even more strongly. You don't want to OCR financial documents with an unverified community script. You want an officially maintained, enterprise-tested capability.

This is the approach behind iflytek/iFly-Skills — an official skill collection covering speech recognition, OCR, translation, proofreading, and multimodal capabilities. Not community-sourced prompt templates, but officially maintained capabilities from a team that has been building these specific technologies for years.

iFly-Skills demo

The Composition: Skills + Orchestration

Skills alone aren't enough. A coding skill pack + a coding agent = a working coding workflow.

Similarly, a multimodal skill pack + an orchestration platform = a working office workflow.

This is where iflytek/astron-agent comes in — an enterprise-grade agentic workflow platform for building SuperAgents. It handles the orchestration: "OCR these invoices → extract amounts → translate the contract → proofread the summary → archive everything."

PrimeIntellect-ai/prime-agent (today's #1 with 2,483 stars) proves that developers want agents that "run long and self-improve." But long-running coding agents and long-running office agents need different things — one needs refactoring skills, the other needs OCR skills.

The Takeaway

The skill pack ecosystem is maturing — but it's lopsided. Coding skills have three trending repos and 300K+ combined stars. Multimodal skills have zero.

If you're building office agents, not coding agents, your skill layer needs to look different. And if you're packaging multimodal capabilities for agents, the bar is higher — but so is the opportunity.

Links:

Top comments (0)