What The Viral Screenshots Miss
If you've been on AI-coding X this week, you've seen the screenshots: 132 agents, dozens of commands, 25+ MCP servers all crammed into one Claude Code session. They look impressive. But the install data tells a different story: the developers actually shipping with Claude Code run something much smaller.
Skillselion, a directory that ranks Claude Code skills and MCP servers by real install counts, tracks 67,052 skills and 8,120 MCP servers across 87,803 listings with over 128 million recorded installs. The most-installed items aren't the exotic ones. They're the boring, reliable ones that solve one problem well.
The Minimum Viable Dev Team
You don't need 132 agents. You need four skills and two MCP servers:
Orchestration spine:
using-superpowers(169,910 installs) — teaches Claude when to reach for other skillswriting-plans(170,580 installs) — forces structured planning before execution
Parallel execution:
subagent-driven-development(134,040 installs) — fans independent steps to subagents so big tasks run in parallel instead of one long serial context
Specialist roles:
requesting-code-review(154,036 installs) — catches bugs before they shipsystematic-debugging(172,160 installs) — disciplined bug hunting instead of guess-and-check
That's it. Four skills. Add a fifth (receiving-code-review at 126,571 installs) if you want the full review loop. Then add two or three MCP servers for your real stack — a database server, a browser server, your API.
Why This Works
The key insight: these skills change how Claude Code works, not just what it can call. writing-plans prevents context drift on long tasks. subagent-driven-development turns serial prompts into parallel work. Together, they make Claude Code behave like a small team without the overhead of 132 separate agent definitions.
The Security Trap Everyone Misses
The viral threads skip a critical detail: a "reviewer" agent that still holds a Bash tool is not read-only. A Bash tool can still run commands that change the repo, even with Write denied. If you want a genuinely read-only role, gate it with a PreToolUse hook, not just a tools allowlist.
How To Install This Stack
# Install the four core skills
claude code skill install using-superpowers
claude code skill install writing-plans
claude code skill install subagent-driven-development
claude code skill install requesting-code-review
# Add systematic-debugging for bug work
claude code skill install systematic-debugging
# Add MCP servers for your stack (example: database + browser)
claude code mcp install @modelcontextprotocol/server-postgres
claude code mcp install @anthropic/server-playwright
Then add roles only when a task actually needs them. Not because a screenshot had 25 MCPs.
When To Expand
Add more skills when you hit a specific wall: webapp-testing (109,599 installs) when you need automated browser testing, executing-plans (141,389 installs) when plans start drifting during execution. But start with the minimum. The developers with the most installs aren't the ones with the longest config files.
Source: dev.to
