Two of the top three projects in this daily ranking are really the same kind of find: text instructions that make your AI coding agent behave better before you touch the codebase. DietrichGebert/ponytail, at the top with 140 star events, tells the agent to look for existing code it can reuse before writing anything new, and to keep validation and error handling intact instead of stripping them out. Try it when you have a small fix that your agent keeps turning into a framework. mattpocock/skills focuses on understanding the task: one of its editable skills makes the agent interview you with detailed questions before it starts typing, which helps when the problem is a requirement that was never actually said out loud.
If you pick between them, ask yourself what usually goes wrong. When the agent adds too much, ponytail’s restraint is the better fit. When it builds the wrong thing confidently, mattpocock/skills’ question-first workflow will catch that earlier. Both are meant to be edited, so read the rules once and rewrite them in your own words. Further down the ranking, chopratejas/headroom compresses logs and tool outputs before they reach the model, worth a try if long repetitive inputs are eating your context window, but run it against a real task to confirm the details you care about survive.