this is all f*cking free
the 9 most-starred Claude Code setups on GitHub have 1,877,443 stars between them, and each one fixes a different bad habit of your agent
find your agent's worst habit, then grab the fix:
it starts coding before anyone agreed on a plan
superpowers, 305,441 stars - brainstorm, plan, tests and review as mandatory steps
↳
github.com/obra/superpowers
it ships code nobody tested
mattpocock/skills, 303,823 - /tdd, /diagnosing-bugs and /code-review from a working engineer
↳
github.com/mattpocock/skills
it runs with no guardrails
ECC, 289,420 - skills, instincts, memory and security in one harness
↳
github.com/affaan-m/ECC
it writes 1,000 lines when 100 would do
andrej-karpathy-skills, 224,286 - one CLAUDE. md built on Karpathy's list of LLM coding pitfalls
↳
github.com/multica-ai/andrej…
it can't make a real Word file or spreadsheet
anthropics/skills, 183,921 - the docx, xlsx, pptx and pdf skills behind Claude's own files
↳
github.com/anthropics/skills
it writes code that didn't need to exist
ponytail, 176,809 - in its own Opus 5.5 benchmark: 53% less code, 41% less time
↳
github.com/DietrichGebert/po…
its UI looks like every other AI app
ui-ux-pro-max, 140,742 - generates a full design system for your project first
↳
github.com/nextlevelbuilder/…
it works alone when you need a team
gstack, 137,792 - Garry Tan's setup: CEO, eng manager, designer, QA and release roles as slash commands
↳
github.com/garrytan/gstack
it burns tokens on chatter
caveman, 115,209 - says less and reads less: 33.2% fewer input tokens across 54 Claude Code runs
↳
github.com/JuliusBrussee/cav…
start with the one that matches last week's worst session. install all nine at once and your setup turns into a mess
67 more skills that turn the same $20 plan into a dev team are in the full list below ↓