Skill Creator

Skill Creator

A skill is a reusable, opinionated expert guide that makes an AI agent dramatically better at a specific domain. Not a prompt. Not a template. A skill is a production-grade system that encodes deep domain expertise, principled workflows, concrete examples, and an unambiguous definition of done — so the agent executes it the same way every time, without the user having to re-explain.

Your job in this skill is to help the user go from "I want a skill for X" to a skill that works reliably in the real world.

Skill Creator Principles

• Understand before writing. A skill written without understanding the user's intent produces generic instructions that could apply to anything. Ask the right questions first.
• Skills generalize — test cases are just accelerants. You iterate on 3–5 examples because it's fast. But the skill needs to work on the 10,000th prompt that nobody has seen yet. Avoid overfitting rules to the specific examples.
• Explain the why, not just the what. Instructions that say "always do X" without explanation get followed blindly in the wrong context. Instructions that say "do X because Y" get applied with judgment. Smart models respond better to reasoning than to commandments.
• Lean beats padded. Remove every instruction that isn't actively improving outputs. A 600-line skill where 200 lines are filler produces worse results than a tight 300-line skill where every line earns its place. If you find yourself writing ALWAYS/NEVER in all caps, step back — try explaining the reasoning instead.
• The description field is a trigger, not a summary. The description is what the agent reads to decide whether to load the skill. It must contain explicit trigger conditions: "Use when the user asks to..." — not just what the skill is about.
• Test against reality, not intuition. You cannot know if a skill works without running it. Gut-feel improvements compound errors. Run test cases. Look at the outputs. Iterate on evidence.

The Skill Creation Loop

Understand intent
       ↓
Write draft skill
       ↓
Run test prompts → review outputs
       ↓
Improve skill based on feedback
       ↓
Repeat until quality is satisfactory
       ↓
Optimize description for triggering accuracy
       ↓
Ship

Jump into this loop wherever the user is. If they already have a draft, skip to testing. If they're still fuzzy on intent, start at understanding. Always be flexible — if they want to skip evals and iterate by feel, that's a valid choice.