← ClaudeAtlas

write-a-skilllisted

Turn a plain-language description of a desired behavior into a properly structured, installable SKILL.md following Anthropic's skill-authoring best practices: trigger-optimized description, explain-the-why instructions, output templates, examples, and anti-trigger calibration. Use this skill whenever the user says "make a skill", "write a skill for", "turn this into a skill", "create a SKILL.md", describes a behavior they want Claude to follow every time, or complains about re-explaining the same instructions across chats (a recurring-behavior request is a skill request in disguise). Do NOT use for one-off task instructions, MCP server development, or general prompt-writing advice.
innovationengines/skills-repo · ★ 0 · AI & Automation · score 70
Install: claude install-skill innovationengines/skills-repo
# Write a Skill A skill is an onboarding guide for an agent: a folder with a SKILL.md that loads on demand and teaches Claude a repeatable behavior. This meta-skill converts "here's what I want" into that document, correctly structured, so the user doesn't need to know the format, the YAML rules, or the triggering mechanics. The stakes are asymmetric. A skill gets invoked hundreds of times by someone who will never read its source, so an hour of care at authoring time is amortized across every future invocation, and a flaw (a description that never triggers, an instruction that overfits) is paid on every one. ## Step 1: Extract the intent From the user's description, and from the conversation itself if they said "turn THIS into a skill" (mine the thread for the workflow, the corrections they made, the output format they accepted), establish: 1. What behavior should the skill produce? 2. When should it fire: what would the user actually type? 3. When must it NOT fire: the near-miss cases? 4. What does a good output look like: is there a fixed format? 5. Does anything need to be deterministic (a script) rather than generated fresh each time? Ask only for what's genuinely missing, 3-5 questions at most. The questions about anti-triggers (#3) and output format (#4) are the ones users never volunteer and always have opinions about. ## Step 2: Write the frontmatter **Name**: lowercase, hyphens, max 64 chars, matching the folder name. Verb phrases or clear nouns: `fact-ch