Agent Skills

PRB's collection of AI agent skills. Designed to work across agents, but primarily built for Claude Code and Codex.

[!WARNING]
These skills are optimized for my personal setup and workflow. If you install them, do your own due diligence and customize them to fit your stack and agents. No warranties, guarantees, or support are provided — use at your own risk.

Installation

npx skills add PaulRBerg/agent-skills

Skills

Skill	Description
autoresearch	Autonomous experiment loop
biome-js	BiomeJS linting/formatting guidance
bump-deps	Node.js dependency updates
bump-release	Release workflow with changelog, tagging
cli-cast	Foundry cast CLI guidance
cli-gh	GitHub CLI operations
cli-just	Just command runner guidance
code-polish	Combined simplification and review
code-review	Expert code review
code-simplify	Code simplification and refactoring
coderabbit	CodeRabbit PR comment analysis and local review triage
commit	Git commit with conventional commits
create-skill	Bootstrap a new agent skill
delayed-command	Wait and execute bash command
diagram-monochrome	Monochrome technical schematic diagrams
effect-ts	Effect-TS patterns and guidance
evm-chains	EVM chain metadata and RouteMesh reference
git-squash	Squash PR branch with semantic commit message
md-docs	Markdown documentation management
oracle-codex	Codex oracle for planning
process-cleanup	Detect and reap zombie processes with procs
tailwind-css	Tailwind CSS v4 styling guidance
vitest	Vitest v4 testing guidance
work	End-to-end task implementation
yeet	GitHub contribution workflows

SKILL.md Frontmatter Guide

[!NOTE]
Full reference: Claude Code Skills docs

Invocation Control

These fields control who can invoke a skill — the user, Claude, or both:

Field	Type	Default	Effect	Use when…
`user-invocable`	`boolean`	`true`	Controls visibility in the `/` slash-command menu	Set to `false` for background-knowledge skills Claude should auto-load silently
`disable-model-invocation`	`boolean`	`false`	Prevents Claude from auto-loading the skill; removes its description from context	Set to `true` for side-effect workflows you trigger manually (deploy, commit, …)

Combined behavior:

Frontmatter	`/` menu	Claude auto-invokes	Description in context
(defaults)	Yes	Yes	Yes
`disable-model-invocation: true`	Yes	No	No
`user-invocable: false`	No	Yes	Yes
Both `true` / `false`	No	No	No (skill is dead)

Execution Context

The context field controls where a skill runs:

Value	Behavior
(default)	Runs inline in the current conversation — the skill content is injected into the active context
`fork`	Runs in an isolated subagent — no access to conversation history; skill content becomes the subagent's prompt

When context: fork is set, the optional agent field selects the subagent type:

`agent` value	Description
(default)	`general-purpose` — full read/write tools
`Explore`	Read-only tools optimized for codebase exploration
`Plan`	Read-only tools for designing implementation plans
Custom agent	Any subagent defined in `.claude/agents/`

References

License

MIT