If your agent keeps turning a tiny edit into a tiny framework, click DietrichGebert/ponytail first. The useful idea is blunt: look for code you can reuse before adding anything new, and do not drop validation or error handling to make a diff look smaller. Try it on the last messy pull request where the model invented an abstraction for a one-off task; keep the same prompt and see whether the result is smaller without losing checks.
For a feature that should take several steps, obra/superpowers is the better fit. It asks the assistant to move through requirements, planning, implementation, tests and review with explicit checkpoints, so you can stop before a vague idea becomes hundreds of lines. mattpocock/skills overlaps if the real failure is the model coding before understanding, since one skill turns the conversation into detailed questions first; pick only the pieces you want rather than importing a whole system.
Before installing any of these or other community skills, run NVIDIA/SkillSpector against the package. It looks for suspicious instructions and credential or supply-chain risks, and remembers triaged findings so an update does not reopen old noise. That pairing is practical: ponytail or superpowers can discipline your workflow, while SkillSpector checks the ingredients you bring into it.