Skills and skill-manager

This repository offers two things:

• skills/: reusable Markdown instructions that teach AI agents specialized workflows.
• skill-manager: the supported way to discover and deploy these skills across Claude Code, Codex/OpenAI agents, and Google Antigravity.

Choose a skill below to inspect or adapt its instructions. Choose skill-manager to install skills, keep one source of truth, and manage updates across agent harnesses.

Skill catalog

drafting-commit-message

Draft a concise, imperative commit title and motivation-focused change bullets from staged or unstaged changes.

grill-me

Explore project context, then interview the user one decision at a time until a plan or design has no unresolved branches.

managing-skills

Operate the complete skill-manager CLI conversationally, with explicit safety and confirmation rules.

reviewing-implemented-work-order

Review a work-order implementation against its job, research, plan, repository patterns, security requirements, and test coverage.

reviewing-my-code

Prepare a focused branch review covering correctness, security, performance, and test coverage.

running-as-maestro

Run an agent as an accountable orchestrator that delegates work to subagents, selects appropriate model tiers, and verifies their output.

Install skill-manager

macOS and Linux:

$ curl -fsSL https://raw.githubusercontent.com/sernst/skills/main/clis/skill-manager/install.sh | sh

Windows (PowerShell):

powershell -c "irm https://raw.githubusercontent.com/sernst/skills/main/clis/skill-manager/install.ps1 | iex"

The installers prompt for a destination when interactive. Windows verifies the release against SHA256SUMS; the POSIX installer verifies it when sha256sum or shasum is available and prints a prominent warning if neither exists.

Customize the install directory

--dir and -Dir accept absolute paths, ~/~/... relative to the active home, and other relative paths resolved from the invocation directory. The resolved absolute path is shown before installation.

On macOS and Linux, pass options through the pipe:

$ curl -fsSL https://raw.githubusercontent.com/sernst/skills/main/clis/skill-manager/install.sh | sh -s -- --dir "./tools/bin" --yes --no-modify-path

Because irm | iex cannot receive parameters, use environment variables on Windows:

$env:SKILL_MANAGER_INSTALL_DIR = '.\tools\bin'
powershell -c "irm https://raw.githubusercontent.com/sernst/skills/main/clis/skill-manager/install.ps1 | iex"

Both scripts also support SKILL_MANAGER_VERSION, SKILL_MANAGER_INSTALL_YES=1, and SKILL_MANAGER_NO_MODIFY_PATH=1. For a manual installation, use the latest release and its SHA256SUMS. For agent-assisted installation, use install.skill-manager.md.

Try an interactive workflow

Add this repository's skill collection, preview one deployment, then apply it:

$ skill-manager source add sernst/skills/skills sernst-skills --label "sernst skills"
$ skill-manager load sernst-skills --filter managing-skills --shared --global --dry-run
$ skill-manager load sernst-skills --filter managing-skills --shared --global
$ skill-manager status managing-skills --shared --global

The dry run shows the same plan without deploying the skill. The next command shows the plan again and asks for authorization. Start a new agent session if your harness scans installed skills only at startup; you can then ask it to use $managing-skills.

Mental model and safety

sources -> discovered skills -> target + scope deployments

• A source is a local directory or GitHub repository path containing skill directories.
• A skill is a directory whose root contains SKILL.md.
• A target is an agent harness's skill directory; built-ins are claude, shared, and antigravity.
• A scope is global (under the manager home) or project (under the current project). Project deployments take precedence.

Manager-owned configuration, cache, backups, and locks live beneath ~/.skill-manager/ by default. Preview load, update, copy, and remove with --dry-run. A dry run does not change deployments, though startup storage migration or a required remote-cache refresh can still update manager-owned state.

Go deeper

• Configuration and filesystem safety
• Human CLI reference
• NDJSON and automation contract
• Goal-oriented cheatsheet
• Architecture and development
• Contributing and releases