coding

claude-skills

Canonical source for agentic skills, routines, hooks, commands, and cross-runtime orchestration patterns.

This repo started as agent-orchestration, the WRFCoin sprint infrastructure repo. It is now being evolved into a broader skill ecosystem repo: still grounded in the tmux/persona sprint work that proved the pattern, but organized so skills can be versioned, packaged, audited, installed across devices, and eventually released publicly.

What Lives Here

• claude/ - Claude Code skills, commands, hooks, and routines.
• codex/ - Codex skills and runtime-specific workflow ports.
• gemini/ - Gemini CLI/runtime parity work, when added.
• skills/ - Portable user skills that are not tied to one runtime.
• docs/ - Architecture, versioning, controls, benchmarks, and release notes.
• manifest.yaml - Canonical version registry for every tracked skill.

Operating Model

The repository treats a skill as a versioned product, not a pasted prompt. Every shippable skill should have:

• a SKILL.md with validator-compatible frontmatter (name, description);
• a manifest entry in manifest.yaml carrying the canonical canonical_version and last_updated;
• a changelog entry — either as a per-skill CHANGELOG.md kept outside the validator-checked skill files, or as notes on the manifest entry;
• a tagged release before zipping or upload;
• a lightweight validation or benchmark story;
• a review evidence contract that states changed behavior, validation, known gaps, reviewer focus, and the agent's own merge bar.

Runtime-specific pieces stay under their runtime folders. Shared skills that should work across Claude Desktop, Claude Code, Codex, Codex Desktop, Gemini CLI, and future runtimes live under skills/.

Versioning

Canonical version metadata lives in manifest.yaml:

skills:
  skill-name:
    canonical_version: 1.0.0
    runtime: shared
    repo_path: skills/skill-name
    last_updated: 2026-05-09
    status: active

SKILL.md frontmatter stays minimal (name + description) for compatibility with the bundled skill-creator validator. The validator accepts metadata as a nested key, so a future migration can carry version data inline as metadata.version / metadata.last-updated without waiting on a validator update. Until that migration is decided, manifest.yaml is the only authoritative version source.

Each skill is independently versioned with semver. Tags are namespaced:

instrument-maker/v4.4.6
tmux-v2/v2.0.0
idea-incubator/v1.0.0

See docs/skill-versioning.md. See docs/manifest-drift-checks.md for the sprint-manager and CI smoke commands.

Controls

The next phase of this repo is about controls: trigger accuracy, routing, runtime adapters, issue-backed handoffs, routine support, and version drift detection. The guiding principle is simple: skills should be easy to install and portable, but hard to silently drift.

See docs/skill-controls.md. For the operator workflow that keeps CLI, desktop, laptop, and review-only copies aligned, see docs/cross-device-skill-sync.md.

Review Gates

Skill PRs are reviewed against an evidence-based gate, not just a plausible diff. Use docs/review-gates/agentic-skill.md for skill, command, hook, adapter, and benchmark changes. Use docs/review-gates/pr-evidence-contract.md as the minimum PR-body template for agent-authored work.

Provenance

The first production use case was WRFCoin: a multi-repo blockchain sprint system using named tmux personas, handoff files, GitHub issues, PR review gates, and Claude/Codex/Gemini runtime parity. That history remains in claude/, codex/, and the architecture docs because it is valuable field evidence.

The current expansion generalizes that system for musical instrument design, maker workflows, yoga/class-planning skills, idea incubation, and public skill release.

Status

Private-first while the repo is being scrubbed, reorganized, and benchmarked for public release.