| name | skill-creator |
| description | Collaboratively design and author agent skills with correct frontmatter, naming, and placement conventions. |
Skill Creator
Optional: Quick fit check
Good fit if the user wants to:
- Create a new skill
- Improve an existing
SKILL.md
- Decide where a skill should live (project/global/custom)
Not needed if the user is asking for a normal code change, debugging help, or a one-off question.
If this skill isn’t needed, ignore it and continue.
Outcomes
When used, this skill helps produce:
- A new skill folder containing a valid
SKILL.md
- A clear, discovery-friendly
description
- Optional
references/ and scripts/ only when needed
How skills work (progressive disclosure)
This skill follows the Agent Skills model:
- Discovery: agents load only
name + description for all skills.
- Activation: when this skill is selected, the agent reads
SKILL.md.
- Execution: the agent reads
references/ files or runs scripts/ only when needed.
Design new skills the same way: keep the main file high-signal, and push deep details into files that are read only on the branch that needs them.
Core rules
- Skills are discovered by frontmatter only:
name + description.
- Skill
name must match the directory name and pass naming constraints.
- Default to a single-file skill.
- Add
references/ only when it reduces token load or avoids confusion.
- Add
scripts/ only for deterministic, mechanical, or fragile steps.
Minimal spec constraints (inline)
name:
- 1–64 characters
- lowercase letters/numbers/hyphens only
- no leading/trailing
-, no --
- must match the skill directory name
- Recall:
description is used for discovery. It should say what the skill does and when to use it, and include concrete trigger keywords.
Workflow
1) Gather requirements (collaborative)
This is a collaborative workflow: ask questions first, then draft, then confirm.
Ask for:
- Skill name (kebab-case)
- A single-sentence
description that makes the discovery decision easy
- What the skill should reliably produce (outputs/artifacts)
- What questions the agent must ask up front (inputs)
- Guardrails (what to avoid, what not to assume)
Do not generate files until the user confirms the name, description, and placement.
2) Choose placement (three options)
Ask the user to choose where the new skill should live:
- Project local (Recommended):
.<tool>/skills/<skill-name>/SKILL.md
- Global:
~/.<tool>/skills/<skill-name>/SKILL.md
- Other (absolute path): user provides an absolute path
Rules for Other:
- If the provided path ends with
SKILL.md, treat it as the exact file path.
- Otherwise treat it as a directory path and place the skill at
<path>/<name>/SKILL.md.
- If the path isn’t absolute, ask again for an absolute path.
3) Validate before writing
Before creating files, verify:
name satisfies the constraints above
- The directory name will exactly match
name
description is 1–1024 characters and describes when to use the skill
4) Author SKILL.md
Write the new SKILL.md directly. Use this structure as a starting point (adapt as needed):
---
name: <skill-name>
description: <One sentence describing when this skill should be used>
---
# <Skill Title>
## Collaboration
- Ask the user for missing inputs.
- Summarize intended outputs and file changes.
- Wait for confirmation before creating or editing files.
## Inputs I need
- <question 1>
- <question 2>
## Workflow
1. <Step 1>
2. <Step 2>
3. <Step 3>
## Guardrails
- <constraints>
Guidelines:
- Prefer concise, step-based workflows.
- Use gates/decision points when paths branch.
- Keep the skill focused on what the user’s environment and repo actually require.
4a) References and scripts (relative paths)
Skills are plain SKILL.md files (no template rendering). When referencing bundled files, always use relative paths from the skill root:
- References:
/references/<doc>.md
- Scripts:
/scripts/<script>.mjs (or .ts if needed)
5) Add references/ (branch only)
Default: do not add references/.
Add references/ only when it keeps SKILL.md smaller and more usable (e.g. approaching ~500 lines, domain-specific docs, long examples).
If you decide you need references, first read:
/references/references-guide.md
When linking to bundled reference files from SKILL.md, use a relative path:
See `/references/<doc>.md`.
6) Add scripts/ (branch only, Node-only)
Default: do not add scripts/.
Add scripts only for deterministic, mechanical steps that benefit from automation (validation, scaffolding, transforms).
If you decide you need scripts:
-
Read script conventions:
/references/scripts-overview.md
-
Use .mjs scripts and the Node standard library only (node:* imports).
-
Validate scripts parse before you rely on them:
- Node:
node -c scripts/<script>.mjs
How to write a good description
A strong description makes the discovery decision obvious.
Patterns that work well:
- Start with an action verb: “Design…”, “Generate…”, “Review…”, “Refactor…”, “Draft…”, “Validate…”
- Include the artifact produced: “...a
SKILL.md…”, “...a migration plan…”, “...a checklist…”
- Include discovery triggers: what user phrases or contexts should cause this skill to activate
Examples:
- “Creates a scoped
SKILL.md for consistent API endpoint changes in this repo. Use when adding or modifying API routes or handlers.”
- “Drafts agent skills for repeatable workflows, keeping
SKILL.md concise and using references/ only for deep details.”