Agent Starter Kit
The skills, plugins, and MCP servers worth installing the moment you open a new
agent — bundled into one install. Instead of hunting them down one by one, run the
installer and the agent starts with real capabilities: it can find and install more
skills on demand, drive a browser, reason with guardrails, design clean UI (Arabic
included), review its own code, and automate the browser + GitHub.
What it installs
Tier 1 — Skills (work on 70+ agents via npx skills: Claude Code, Codex, Cursor, …):
| Skill |
What it gives the agent |
find-skills |
Discovers and installs any other skill on demand — the meta-skill |
agent-browser |
A real browser the agent controls itself (open, click, fill, verify) |
karpathy-guidelines |
Behavior rules: no silent assumptions, minimal solutions, verify before acting |
ui-ux-pro-max |
Design-system intelligence — styles, color palettes, font pairings |
caveman |
Blunt, no-fluff answers when you want signal over politeness |
arabic-design |
Clean Arabic in AI-generated designs (fixes the letter-spacing/RTL traps) |
Tier 2 — Plugins / MCPs (Claude Code only):
| Plugin / MCP |
What it gives Claude Code |
frontend-design |
Anthropic's design guidance for polished, non-generic UI |
superpowers |
TDD + subagent orchestration workflow (brainstorm → design → build → review) |
code-review |
/code-review on any branch/PR — scored, actionable review |
playwright (MCP) |
Browser automation & end-to-end testing |
github (MCP) |
Read/manage GitHub repos, PRs, and issues from the agent |
claude-md-management |
Keeps your CLAUDE.md files organized and current |
How to run it
From the skill directory:
bash install.sh # Claude Code — installs everything
bash install.sh --agent cursor # Cursor — installs the Tier 1 skills
bash install.sh --agent codex # Codex — installs the Tier 1 skills
bash install.sh --agent '*' # every detected agent
bash install.sh --skills-only # skip the Claude-Code-only plugins/MCPs
The script is best-effort and safe to re-run: each item installs independently, and
anything it can't install automatically prints the exact manual command to finish it.
When invoked as a skill
If an agent loads this skill to set up an environment: run install.sh with the right
--agent for the current agent. Then confirm what installed and surface any manual
fallback lines the script printed. In Claude Code, remind the user to restart (or run
/plugin) so the new plugins load.
Honest scope note
- Tier 1 skills are genuinely cross-agent —
npx skills targets 70+ agents.
- Tier 2 is Claude Code only. Codex/Cursor have no "plugin" equivalent. Their two
MCP items (
playwright, github) can still be added to Codex/Cursor through that
agent's own MCP config — the script says so when you pass a non-Claude --agent.
1---2name: agent-starter-kit3description: One-command setup for a new AI coding agent. Installs the essential starter skills (find-skills, agent-browser, karpathy-guidelines, ui-ux-pro-max, caveman, arabic-design) across 70+ agents, plus the core Claude Code plugins/MCPs (frontend-design, superpowers, code-review, playwright, github, claude-md-management). Use when the user is setting up a new project or a fresh agent and asks for "the skills/MCPs I should start with", a "starter pack", or "set up my agent".4---56# Agent Starter Kit78The skills, plugins, and MCP servers worth installing the moment you open a new9agent — bundled into one install. Instead of hunting them down one by one, run the10installer and the agent starts with real capabilities: it can find and install more11skills on demand, drive a browser, reason with guardrails, design clean UI (Arabic12included), review its own code, and automate the browser + GitHub.1314## What it installs1516**Tier 1 — Skills (work on 70+ agents via `npx skills`: Claude Code, Codex, Cursor, …):**17| Skill | What it gives the agent |18|---|---|19| `find-skills` | Discovers and installs any other skill on demand — the meta-skill |20| `agent-browser` | A real browser the agent controls itself (open, click, fill, verify) |21| `karpathy-guidelines` | Behavior rules: no silent assumptions, minimal solutions, verify before acting |22| `ui-ux-pro-max` | Design-system intelligence — styles, color palettes, font pairings |23| `caveman` | Blunt, no-fluff answers when you want signal over politeness |24| `arabic-design` | Clean Arabic in AI-generated designs (fixes the letter-spacing/RTL traps) |2526**Tier 2 — Plugins / MCPs (Claude Code only):**27| Plugin / MCP | What it gives Claude Code |28|---|---|29| `frontend-design` | Anthropic's design guidance for polished, non-generic UI |30| `superpowers` | TDD + subagent orchestration workflow (brainstorm → design → build → review) |31| `code-review` | `/code-review` on any branch/PR — scored, actionable review |32| `playwright` (MCP) | Browser automation & end-to-end testing |33| `github` (MCP) | Read/manage GitHub repos, PRs, and issues from the agent |34| `claude-md-management` | Keeps your `CLAUDE.md` files organized and current |3536## How to run it3738From the skill directory:3940```bash41bash install.sh # Claude Code — installs everything42bash install.sh --agent cursor # Cursor — installs the Tier 1 skills43bash install.sh --agent codex # Codex — installs the Tier 1 skills44bash install.sh --agent '*' # every detected agent45bash install.sh --skills-only # skip the Claude-Code-only plugins/MCPs46```4748The script is **best-effort and safe to re-run**: each item installs independently, and49anything it can't install automatically prints the exact manual command to finish it.5051## When invoked as a skill5253If an agent loads this skill to set up an environment: run `install.sh` with the right54`--agent` for the current agent. Then confirm what installed and surface any manual55fallback lines the script printed. In Claude Code, remind the user to restart (or run56`/plugin`) so the new plugins load.5758## Honest scope note5960- **Tier 1 skills are genuinely cross-agent** — `npx skills` targets 70+ agents.61- **Tier 2 is Claude Code only.** Codex/Cursor have no "plugin" equivalent. Their two62 *MCP* items (`playwright`, `github`) can still be added to Codex/Cursor through that63 agent's own MCP config — the script says so when you pass a non-Claude `--agent`.
jamalmohafil