Start with mattpocock/skills if you have ever watched an agent confidently implement the wrong thing. Its standout skill flips the usual order: before writing any code, the assistant asks you detailed questions about the requirement. Pick just that one skill, drop it into your project, and give it a task you would normally get back wrong, like a bug report with a vague expected behavior. You are essentially trading a little extra typing at the start for fewer rewrites at the end.
DietrichGebert/ponytail and affaan-m/ECC offer two more ways to organize the work: reuse rules and a broader engineering routine. Ponytail’s instructions ask it to reuse existing code first and to keep validation and error handling in place, so those checks do not get lost in an overly aggressive cleanup. ECC is broader, covering planning, tests, review and remembering earlier decisions. If your recurring problem is bloat, try ponytail on one focused refactor. If your agent forgets what it decided two sessions ago, browse ECC and adapt the routines that match how your team actually works.
Two diagram projects in this ranking with different audiences. tt-a1i/archify takes an agent’s structured description of a system and renders interactive HTML and SVG architecture diagrams, and it can diff two snapshots so you can show a team exactly which components and connections changed between versions. cathrynlavery/diagram-design has a different emphasis: you hand it an explanation or existing diagram source and it produces graphics that borrow your publication’s colors and fonts, so a flow chart stops looking like a tool’s default export. Use archify when the question is what changed; use diagram-design when the question is whether it looks like it belongs in your article.
lnkiai/m3e-canvas solves the classic mismatch where drawing is easier than describing. Sketch screens with Material-style components in the browser, click through the transitions between them, and let it turn the result into a prompt for a coding assistant. If you have a multi-screen flow that is painful to narrate in chat, this is a sensible first try: sketch three screens and their links, generate the prompt, and see if the implementation finally matches the flow you had in your head.
For a break from agent tooling, open ashemag/human-atlas. It is an interactive 3D anatomy viewer: select a structure, rotate the model and look at how it sits beside the surrounding parts. If a flat illustration leaves you wondering what is behind an organ, the extra viewing angle gives you another way to explore it.