The 1.9 Million-Star Prescription for Claude Code: Don’t Take It All at Once

Mr. Buzzoni listed on X nine of the most-starred Claude Code setups on GitHub, totaling 1,877,443 stars. The number is impressive. More worth looking at is the problem each repo addresses: writing code before the agent is aligned, shipping untested work, having no guardrails, writing 1,000 lines when 100 would do, failing to produce real Word docs and spreadsheets, writing code that doesn’t need to exist, UIs that look mass-produced, going solo when a team is needed, and burning tokens on chit-chat.

Claude Code is Anthropic’s coding agent, and these setups change the way it works through skills, hooks, rules, and commands. Before installing, you need to know what they change.

Nine setups, nine bad habits

1. Write code first, plan later: obra/superpowers, 305,441 stars. Turns brainstorming, planning, testing, and review into mandatory steps. The repo description is clear: the agent first asks what you want to do, then gives a design, then breaks down the tasks. Repo

2. Shipping code nobody tested: mattpocock/skills, 303,823 stars. Provides skills such as /tdd, /diagnosing-bugs, and /code-review; the author says these come from real engineering, not vibe coding. Repo

3. No guardrails: affaan-m/ECC, 289,420 stars. Puts skills, instincts, memory, and security scanning into one harness. The README also warns against stacking installs, because duplicate installs make hooks run twice. Repo

4. Writing 1,000 lines when 100 would do: multica-ai/andrej-karpathy-skills, 224,286 stars. A single CLAUDE.md based on the LLM coding pitfalls Karpathy listed. Repo

5. Can’t produce real docx, xlsx, pptx, or pdf files: anthropics/skills, 183,921 stars. The same set of skills Claude uses to generate files itself. Repo

6. Writing code that doesn’t need to exist: DietrichGebert/ponytail, 176,809 stars. In the repo’s own Opus 5.5 benchmark, it produced 53% less code and took 41% less time. Repo

7. UI looks like every AI app: nextlevelbuilder/ui-ux-pro-max-skill, 140,742 stars. Generates a complete design system first. Repo

8. Going solo when you need a team: garrytan/gstack, 137,792 stars. Turns the roles of CEO, engineering manager, designer, QA, and release into slash commands. Repo

9. Burning tokens on chit-chat: JuliusBrussee/caveman, 115,209 stars. Talk less, read less; across 54 Claude Code runs, input tokens were 33.2% lower. Repo

Stars are not an audit

In a reply, user Marius du Preez poured cold water on it: stars aren’t an audit—how many people read the hooks before starring and then running them with their own keys? That sentence is worth pausing on. The hooks in these setups can run shell commands, MCP can access credentials, and project instructions go into context. ECC’s README also treats hooks, MCP, and project instructions as executable configuration. Before installing, at least read the hooks and permissions.

Buzzoni’s own advice is to start with the one that matches your worst session from last week. Install all nine at once and your setup will become a tangled mess. User Noaa wanted to add one more: make the agent read the current docs before guessing an API. Shruthink asked which setup actually changed daily workflows. That question is more useful than star counts.

Claude Code setups change the agent’s workflow and permission boundaries. Pick one bad habit, read its hooks, install it, and run it for a week. Then decide on the next one. Star counts are someone else’s result, not your audit report.

The full original list is in Mr. Buzzoni’s X article.