Installs into .claude/skills of the current project.
Are you the author of Commands?
Add the live security badge to your README. It updates with every re-scan.
[](https://www.skillsdirectory.com/skills/eyelock-commands)
---
name: new-skill
description: Scaffold a skill in the current harness, write it properly, and prove it lands
---
Add a skill to the harness in the current directory, then verify it reaches the
vendor rather than assuming it did.
## 1. Scaffold
```bash
ynd create skill <name>
```
That writes `skills/<name>/SKILL.md` with placeholder frontmatter. The `name`
must match the directory.
## 2. Write it
Ask what workflow this automates before writing anything. A skill is a workflow,
not a description — every step should be a command to run, a file to open, or a
pattern to look for.
Replace the placeholder `description` first. It is what the vendor routes on, so
it must say **what the skill does and when to use it**, not just what it is.
Follow `rules/artifact-authoring.md`. If the body grows past a screen, move
detail into `skills/<name>/references/` — those ship with the skill, so anything
the skill tells the reader to open must live there rather than elsewhere in the
repo.
## 3. Prove it
```bash
ynd validate .
ynd lint skills/<name>
ynd preview . -v claude
```
The preview is the step people skip. **Confirm the new file appears in it** — a
skill that validates but does not assemble is not installed, and nothing else
tells you.
If the harness is installed, reload it:
```bash
ynh install .
```