If you have watched an agent confidently start coding something you never asked for, mattpocock/skills is worth opening first. It ranked first with 942 star events, and its most interesting entry is a workflow that makes the assistant interview you about the requirement before touching code. Try it on your next real task: pick one skill, drop it into your assistant, and see whether the questions it asks catch a misunderstanding you would otherwise fix after the fact. DietrichGebert/ponytail tackles the same overeagerness from another angle, with rules that push the agent to reuse existing code and keep validation and error handling intact. For a small fix, that can mean the difference between a tidy patch and a freshly invented framework.
If you are still deciding how a screen should work, sketch it before asking for code. lnkiai/m3e-canvas lets you sketch screens in the browser and test how pages link together, then turns the sketch into a prompt for a coding assistant. A sensible first experiment: draw a three-screen flow for something you are building anyway, click through the transitions, and hand the generated prompt to your agent. If you need to explain a whole system, tt-a1i/archify converts a structured description into interactive diagrams and can compare two snapshots, so a design review becomes ‘here is what changed’ instead of a long verbal tour. For polished figures that match an article’s style, cathrynlavery/diagram-design offers another option, applying your colors and fonts to HTML and SVG output.