reasoning-framework

llm-skills

A multi-format LLM skills marketplace. One portable source compiles to Claude Code Skills, Claude Code Plugins, OpenAI GPTs/Apps, and a vendor-neutral portable format.

Repository: https://github.com/ravenoak/llm-skills

Status

Early release. The first shipping skill is reasoning-framework; more skills will land against the same spec and pipeline.

Install

The same source compiles to several runtimes — pick the path that matches yours. Full instructions for every target are in docs/installing.md.

Claude Code (marketplace, recommended)

In an interactive Claude Code session:

/plugin marketplace add ravenoak/llm-skills
/plugin install reasoning-framework@ravenoak-llm-skills

Each skill is its own installable plugin, with its own version. That means skills ship on independent cadences — reasoning-framework can bump 0.2 → 0.3 without forcing the others to rev (and vice versa); installs of one don't pin the others. Pick what you want; nothing is all-or-nothing. /plugin marketplace update ravenoak-llm-skills picks up new releases. Releases are tagged v*.*.*; the latest tag is the source of truth.

Claude Code (single SKILL.md, no plugin)

Drop the committed SKILL.md into your project's or user-wide skills directory:

mkdir -p ~/.claude/skills/reasoning-framework
curl -L \
  https://raw.githubusercontent.com/ravenoak/llm-skills/main/plugins/reasoning-framework/skills/reasoning-framework/SKILL.md \
  -o ~/.claude/skills/reasoning-framework/SKILL.md

OpenAI custom GPTs / Apps (untested)

For each skill that opts into the openai-gpt target, dist/openai/<id>/ ships a manifest.json and instructions.md. Paste them into a custom GPT's configuration. See docs/installing.md. Status: artifacts are emitted by the build but have not yet been confirmed end-to-end against the OpenAI GPT builder.

Portable format (untested)

The vendor-neutral artifact lives at dist/portable/<id>/skill.json; it validates against spec/skill.schema.json. Status: schema validates, but no external consumer has been tested.

Using a skill

Once installed, the skill activates from its description (the trigger phrase Claude reads). Each skill's SKILL.md documents what it's for, when to use it, and when to skip it. For example, reasoning-framework activates when stress-testing a conclusion or auditing a complex claim, and is meant to be skipped for time-pressured decisions or mechanical lookups.

Local development

brew install go-task        # or: go install github.com/go-task/task/v3/cmd/task@latest
task setup                  # installs Node + Python tooling
task check                  # runs CI gate locally

Layout

• spec/ — the canonical portable skill format (JSON Schema)
• plugins/<id>/ — per-skill plugin source + committed Claude artifacts (each is a standalone Claude Code plugin)
• tools/ts/ — skillsmith CLI (npm)
• tools/py/ — llm-skills CLI (PyPI)
• dist/ — build outputs for OpenAI and portable targets (gitignored)

Documentation

• docs/installing.md — install and update instructions for every target
• docs/authoring.md — write a new skill
• docs/targets/ — per-target build details (Claude skill, Claude plugin, OpenAI GPT)
• docs/governance.md — release and contribution policy
• docs/superpowers/specs/2026-05-21-llm-skills-host-design.md — original design spec

License

MIT — see LICENSE.