Skip to content

[RFC]: Skillware tiered architecture — soft skills, hard skills, and ultimate skills #60

@rosspeili

Description

@rosspeili

Summary

Propose a formal tiered skill architecture for Skillware that distinguishes between three classes of skills based on implementation complexity, dependency requirements, and deployment context.

Motivation

As the Skillware registry grows, a single flat skill standard is increasingly difficult to apply uniformly. A prompt-based reasoning skill has fundamentally different requirements than a skill that calls external APIs or runs cloud-only enterprise pipelines. The current standard conflates these cases, leading to inconsistency in manifests, documentation, and contributor expectations.

A tiered architecture would:

  • Set clear expectations for contributors about what level of complexity their skill represents.
  • Allow the loader and documentation to handle each tier differently where appropriate.
  • Create a natural progression from simple to advanced skills that mirrors how the ecosystem can grow.

Detailed Design

Tier 1 — Soft Skills (Anyone can contribute)

Prompt-based skills with no external dependencies. The skill's logic lives entirely in instructions.md. There is no skill.py or only a trivial passthrough implementation. These are analogous to the skills.md by Anthropic, Google, etc.

  • No external API calls.
  • No additional Python dependencies beyond the core framework.
  • Can be run with any supported LLM provider without configuration.
  • Suitable for contribution by anyone familiar with prompt engineering.

Tier 2 — Hard Skills (Anyone can contribute)

The current Skillware standard. A full implementation in skill.py with a manifest.yaml, card.json, and instructions.md. May call external APIs. Requires standard Python dependencies. Runs locally with appropriate environment variable configuration.

  • May have external dependencies declared in the manifest.
  • Requires environment variable configuration for API keys.
  • Tested with the standard CI pipeline.

Tier 3 — Ultimate Skills (ARPA Core Enterprise Solutions Devs Only)

Enterprise-grade skills that may require cloud infrastructure, paid API access, proprietary data sources, or managed deployment. Not intended for local self-hosting without configuration.

  • May require cloud provider credentials or KMS integration.
  • May depend on private or paid data sources.
  • Documentation directs users to contact the maintainers for deployment support.
  • Not expected to pass the standard CI suite without additional configuration.

Proposed manifest change

skill_tier: soft | hard | ultimate

### Open questions
- Should tier be enforced by CI or only by convention?
- Should soft skills have their own directory separate from skills/?
- Should the registry surface tier information on skillware.site?
- Other?

### Drawbacks

- Adds complexity to the manifest schema and documentation.
- The line between tiers may be subjective in some cases.

Metadata

Metadata

Assignees

No one assigned

    Labels

    core frameworkChanges to loader, env, or base classes.discussionOpen discussion for RFCs and proposals.enhancementNew feature or request
    No fields configured for Feature.

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions