Each skill is a plain folder with a SKILL.md — drop it into your project and the agent starts using it immediately. No daemon, no runtime, no build step.
curl -fsSL https://raw.githubusercontent.com/mifunedev/skills/master/scripts/install.sh | bash -s -- install <skill-name> --scope project| Flag | Default | Purpose |
|---|---|---|
--scope project |
project |
Install into the current git repo |
--client agents|claude|harness |
agents |
Target destination (.agents/skills/, .claude/skills/, or both) |
The installer pins the registry commit SHA in .mifune/skills.lock — re-installs of the same version are deterministic.
11 skills across 4 categories:
| Skill | Category | Description |
|---|---|---|
agent-browser |
dev-workflow | Open a URL in the headless browser with preflight health check |
ci-status |
dev-workflow | Poll CI after push; reports pass/fail with failure details |
prd |
dev-workflow | Generate a Product Requirements Document for a new feature |
release |
dev-workflow | Cut a CalVer release: version, tag, push, CI poll, verify image |
ste |
dev-workflow | Write technical prose in Simplified Technical English, with a deterministic checker |
worktrees |
dev-workflow | Manage .worktrees/ lifecycle: create, list, remove, clean, audit |
delegate |
orchestration | Parallel execution coordinator — decomposes plans into wave-executed sub-agents |
system-one |
integration | Typed choice, yes/no, and score judgments from System One: Jev by default, self-hosted Laya as the alternative |
post-bridge |
integration | Publish posts, upload media, and schedule content via the Post Bridge API |
remote-sandbox |
agro | Provider-generic VM driver, plus the AGRO fresh-install check and hosting matrix on exe.dev and Vercel Sandbox |
interview |
agro | Adaptive pre-work clarifier — batches 2–4 task-specific questions, then proceeds |
The canonical index — versions, checksums, and requires-tools per skill — is registry.json.
Adding a new skill takes under 10 minutes.
- Create the folder —
mkdir skills/<name>(lowercase, hyphens, ≤ 64 chars) - Write the skill — start from an existing
SKILL.md; fill inname,description, and the imperative instruction body - Add a license — copy the root
LICENSEintoskills/<name>/LICENSE - Recompute checksums —
./scripts/refresh-checksums.sh - Register the skill — add an entry to
registry.jsonwithname,path,version,description,category,requires-tools,clients,license,added,updated - Validate —
./scripts/validate.shmust reportPASS - Commit —
skills/<name>/andregistry.jsonin the same commit
Checksum algorithm: docs/checksum.md. Portability rules (Claude Code-specific keys go under metadata.mifune.claude-code.*): docs/portability.md.
| Path | Purpose |
|---|---|
registry.json |
Canonical index — one entry per skill with version, checksum, and category |
skills/<name>/ |
One subfolder per skill (SKILL.md + per-skill LICENSE) |
scripts/install.sh |
Bash installer — curl | bash entry point |
scripts/validate.sh |
CI gatekeeper — 19 checks: schema, registry parity, checksum integrity |
scripts/refresh-checksums.sh |
Recompute and write checksums back into registry.json |
scripts/test-install.sh |
Hermetic test harness for install.sh (6 scenarios) |
docs/ |
Checksum algorithm, portability policy, schema |
template/ |
Skill boilerplate for mifune skills new <name> (V1) |
| Version | Status | Highlights |
|---|---|---|
| V0 | ✅ Current | 11 skills, hand-written registry.json, Bash installer, CI validator |
| V1 | 🔵 Planned | TypeScript CLI (@mifune/skills-cli), npm + GHCR distribution, skills.mifune.dev catalog |
| V2 | 🔵 Planned | Sigstore signing, oh skills wrapper, standalone binary |
| V3 | 🔵 Planned | Community contributions, federated registries |
Full design notes: ryaneggz/open-harness PR #306.
This repository is licensed under MIT.
Each skill in skills/<name>/ carries its own MIT LICENSE file as a per-skill override — an explicit affordance of the Agent Skills specification.