Build an agent skill with AI
Turn a repeated correction into a reusable instruction. Compare a fresh agent’s first attempt with its work after the skill is installed.
Build an agent skill with AI when you can point to a recurring mistake. For example, your agent may write a perfectly readable meeting summary yet quietly promote suggestions into decisions. A brief instruction about what counts as agreed is more useful here than a whole guide to taking minutes.
With skill-shaper, a fresh agent tries a safe example before the instructions are written. Its first attempt gives you a reference for judging the next one. If the agent already handles the case well, you may need nothing new. If it selects the wrong existing skill, an edit to that skill’s description may be enough.
What it does
- Decide whether a recurring problem needs a skill or a project rule
- Test which wording selects an existing skill
- Separate two workflows that share the same vocabulary
- Keep project conventions in the right directory
- Produce examples that should and should not activate the skill
Install and use
- Use a skill-compatible agent with access to the appropriate folder. Choose a harmless task that another session can repeat with the same tools.
- Ask: “Make a skill for recording agreed decisions. First try these meeting notes without the skill and show me which suggestions were recorded as decisions.”
- Try the resulting instructions on a second example and a near miss, such as a request to brainstorm proposals. Keep the rule if it improves the intended case without catching the other one.
FAQ
Source
Made by overment. Original skill: skill-shaper, published on GitHub under the MIT licence.
View on GitHub