Nick Perry shares three practices for building Claude skills from an Anthropic engineer. Keep a "darn it" list of every mistake Claude makes to improve iteratively. Organize skill assets in one folder so Claude pulls only what it needs, keeping context clean. Give Claude context and constraints instead of rigid instructions, letting it adapt rather than railroading the model.
Here are 3 tips for building Claude skills that an engineer from Claude just dropped. Quick head up, they're some of the best I've ever seen. First, build a "darn it" section. Because the most important part of a skill isn't the instructions, it's the list of every single thing Claude keeps getting wrong. So now every time it messes up, you add it to that list. And that's actually how the skill gets better over time. Second, use the file system. See, a skill isn't just one markdown file. It's a whole folder for Python scripts, references, examples, and templates. Claude only pulls them in when it actually needs them, so they never clutter up your context. And finally, stop railroading it. If your instructions are too rigid, Claude can't actually work. So give it the context and the constraints, then let it adapt to the real situation.
