claude-skills
paologum/claude-skills
My Claude Code plugin — reusable skills and subagents.
Instalação
npx convez-ai paologum/claude-skills1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
README
A tradução para o português deste README ainda está na fila.
claude-skills
A personal collection of Claude Code skills and subagents, packaged as an installable plugin.
What's in here
| Type | Name | What it does |
|---|---|---|
| Skill | /pr-description | Generates a concise PR description for the current branch (Why / What / Tested / Demo). Auto-triggers when you ask Claude to draft / write / open a PR. |
| Skill | /unity-local-multiplayer | Sets up Unity Multiplayer Play Mode (MPPM) in the current project for up to 4 local virtual players. Adds the package, writes a Player-Tag-aware bootstrap that auto-starts Host vs Client (Mirror or Netcode), documents the Editor steps. |
| Skill | /unity-test | Runs Unity Test Framework tests (EditMode/PlayMode) via batch CLI. Auto-detects Unity version, parses NUnit XML, reports pass/fail with failure details. |
| Skill | /unity-build | Builds the Unity project for Windows / Mac / WebGL / iOS / Android / Linux. Creates Assets/Editor/BuildScript.cs if missing. |
| Skill | /unity-new-project | Scaffolds a new empty Unity project, patches Packages/manifest.json with URP/HDRP/2D + Input System + Test Framework, writes a Unity .gitignore. |
| Skill | /coding-principles | Reference card of C# / Unity / general engineering principles. Auto-loaded into the coder agent; also invokable directly as a style-guide lookup. |
| Agent | coder | Executes a coding plan from a planner. Strictly follows the principles, refuses to expand scope or add abstractions the plan didn't ask for. Unity-aware: writes C# and Editor builder scripts for scene/UI changes; refuses to hand-edit .unity / .prefab YAML hierarchies. |
| Agent | researcher | Performs disciplined web research with source-cited reporting. Prefers primary sources, tags every claim with [Source] / [Inference] / [Conflict] / [Gap], and refuses to assert facts it cannot cite. |
Install
Inside any Claude Code session, run:
/plugin marketplace add paologum/claude-skills
/plugin install claude-skills@claude-skills
/reload-plugins
That's it. Plugins persist across new sessions on the same machine.
To update later:
/plugin marketplace update claude-skills
/reload-plugins
How to use
Skills are auto-invoked when Claude recognizes your request — no slash needed:
"Write a PR description for this branch" → auto-triggers
/pr-description
Or invoke manually:
/pr-description # uses the default base branch
/pr-description develop # against a specific base branch
Adding a new skill
1. Create the file
skills/<your-skill-name>/SKILL.md
The directory name becomes the slash command (/your-skill-name).
2. Write the frontmatter
---
name: your-skill-name
description: One sentence on what it does + when Claude should auto-invoke it. Include phrases the user would naturally say.
allowed-tools: "Read Bash(git *) Grep"
argument-hint: "[optional-arg]"
---
The description is the single most important field. It's how Claude decides whether to auto-invoke. Write it so it includes the natural trigger phrases — e.g., "Use when the user asks to X, Y, or Z."
Common frontmatter fields:
| Field | Purpose |
|---|---|
name | Defaults to directory name. |
description | When Claude should use this skill. Required for auto-invocation. |
allowed-tools | Pre-approves tools — "Read Bash(git *) Grep" etc. |
argument-hint | Autocomplete hint shown in the / menu. |
model | sonnet, opus, haiku, or inherit. |
disable-model-invocation | true = manual-only (no auto-trigger). |
3. Write the body
The markdown below the frontmatter is the instructions Claude follows when the skill runs.
Use the !`shell command` syntax to inject live context that runs before Claude sees the skill:
## Current diff
!`git diff HEAD`
## Task
Summarize the diff above...
See skills/pr-description/SKILL.md for a complete example.
4. Test it locally
Without committing, you can test against a live session:
claude --plugin-dir /Users/paologum/github
This loads your in-progress changes for that one session. Edit SKILL.md, then in the session run /reload-plugins to pick up changes.
5. Commit and push
git add skills/<your-skill-name>
git commit -m "Add /<your-skill-name> skill"
git push
Anyone with the plugin installed will get it after running /plugin marketplace update claude-skills + /reload-plugins.
Adding a new agent
Agents are subagents Claude delegates to via the Agent tool — useful for parallel work, isolated context, or specialized behavior.
1. Create the file
agents/<your-agent-name>.md
(Single file, not a directory.)
2. Write the frontmatter
---
name: your-agent-name
description: When Claude should delegate to this agent.
tools: Read, Grep, Glob, Bash
model: sonnet
---
Common fields:
| Field | Purpose |
|---|---|
name | Required. Lowercase, hyphens. |
description | Required. When Claude should delegate. |
tools | Allowed tools. Omit to inherit all. |
model | sonnet, opus, haiku, inherit. |
permissionMode | default, acceptEdits, plan, etc. |
3. Write the system prompt
The markdown body is the agent's system prompt. Be explicit about behavior, output format, and constraints.
4. Document, commit, push
Add a row to the table at the top of this README, then commit and push.
Releasing changes (author side)
Every push that changes a skill or agent must bump the plugin version — otherwise installed clients silently miss the update (their cache is keyed by version).
# 1. Edit the skill / agent / docs
$EDITOR skills/your-thing/SKILL.md
# 2. Bump the version in .claude-plugin/plugin.json
# - patch (0.2.0 → 0.2.1): edits to existing files
# - minor (0.2.0 → 0.3.0): new skill/agent or behavior change
# - major (0.2.0 → 1.0.0): breaking change to an existing skill
# 3. Update the catalog table at the top of this README (only when adding/removing)
# 4. Commit and push
git add -A
git commit -m "Describe the change"
git push
That's it from the author side. No tag, no release process, no marketplace submission — the marketplace.json in this repo points at main.
Pulling updates (user side)
In any Claude Code session:
/plugin marketplace update claude-skills
/reload-plugins
That should be enough — the marketplace catalog refresh sees the new version, and reload pulls the matching cache.
If the new content doesn't show up (the cache is stuck at the old version), do a forced reinstall:
/plugin uninstall claude-skills
/plugin install claude-skills@claude-skills
/reload-plugins
Verifying an update worked
The reload counter (Reloaded: 1 plugin · 0 skills · 7 agents) is misleading — don't trust it. Use one of these instead:
Shell — authoritative ground truth:
ls ~/.claude/plugins/cache/claude-skills/claude-skills/*/{skills,agents} 2>/dev/null
Shows the version directory and every skill / agent the cache contains. If it matches what's on main, you're up to date.
In-session:
The skill picker (/ menu) and the <system-reminder> available-skills list include every loaded skill, namespaced as claude-skills:<name>. Six of those = all skills loaded. The coder agent shows up when you ask Claude to delegate to it.
Troubleshooting
| Symptom | Fix |
|---|---|
| Skill doesn't auto-trigger | Rephrase the request to match the description. If it still doesn't trigger, broaden the description with more trigger phrases. |
Edits to a SKILL.md not taking effect | Run /reload-plugins. New skill directories may require a session restart. |
/plugin install says "not found" | Run /plugin marketplace update claude-skills first to refresh the catalog. |
| Skill runs but a tool is blocked | Add it to allowed-tools in the frontmatter. |
Repo layout
.claude-plugin/
├── plugin.json # plugin manifest
└── marketplace.json # makes this repo its own marketplace
skills/
└── <name>/SKILL.md # one directory per skill
agents/
└── <name>.md # one file per agent
CLAUDE.md # context for Claude when working in this repo
README.md # this file