If your AI coding agent keeps turning a small fix into a new framework, DietrichGebert/ponytail is worth a look. It is a set of instructions you hand the agent before it codes: check whether existing code can be reused, question whether a new abstraction is actually needed, and keep validation and error handling instead of quietly dropping them. A concrete first try is to pick a task that normally produces a sprawling diff, add the file to your agent’s context, and see if the change stays smaller this time. It leads today with 160 star events, well ahead of the rest of the list.
For another common problem, an agent that starts coding before it understands you, mattpocock/skills includes a skill that asks you detailed questions first. The pieces are short and editable, so choose one and rewrite it in your own words rather than importing everything. If your work means gathering material across several sites, Panniantong/Agent-Reach gives command-line agents tools to read web pages, video platforms and social sources, plus channel diagnostics that show which connections work and which need configuration. A sensible first run is a simple research question that makes you test each channel one by one.