| name | skills-creator |
| description | Scaffold a new agentskills.io SKILL.md from a short description, with the right frontmatter, a focused scope, and a quality checklist |
| state | active |
Skills creator
Use this skill when the user wants to turn a repeatable procedure into a new
agent skill. Your job is to produce a single, well-scoped SKILL.md that follows
the agentskills.io format and the quality
bar set by the network-state-query reference skill in this registry.
Before you write anything
Ask the user for, or infer from context, these four things. Do not skip this
step. A skill with a fuzzy scope is worse than no skill.
- Trigger. In one sentence, when should an agent reach for this skill?
- Scope. What is explicitly in, and what is explicitly out? A good skill
does one job. If you find yourself listing three unrelated jobs, that is three
skills.
- Inputs and outputs. What does the skill need from the user, and what
structured result does it produce?
- Failure modes. What should the agent do when a precondition is not met
(tool missing, device unreachable, ambiguous request)? Name the behavior.
If the answer to "should this be a skill?" is unclear, say so. A one-off task
that will not recur does not need a skill. Recognizing over-engineering early is
itself part of the discipline.
Write the SKILL.md
Produce a directory <name>/SKILL.md with this shape:
---
name: <kebab-case-name>
description: <one line, starts with a verb, says when to use it>
state: draft
---
# <Title>
<One paragraph: what this skill does and when to use it.>
## Steps
<Numbered, deterministic procedure. Each step is an instruction to the agent.>
## Output
<The exact structure the agent should return.>
## Failure modes
<What to do when each precondition fails.>
Rules:
name is kebab-case and matches the directory name.
description starts with a verb and states the trigger, because Gridctl serves
it as the prompt's one-line summary and that is how the user picks the skill.
- Set
state: draft on creation. The author activates it deliberately with
gridctl activate <name> once it has been reviewed. Choosing what loads into
your agent is part of the lesson.
- Least privilege: name only the tools the skill actually needs. Do not invite
the agent to reach for destructive operations.
- Graceful failure: every skill says what to do when a precondition fails.
After writing
- Save to
skills/<name>/SKILL.md (core, curated) or
showcase/skills/<name>/SKILL.md (community floor), then load it into the
registry (./scripts/load-skills.sh or gridctl skill add).
- Validate:
gridctl skill validate <name>.
- Activate when ready:
gridctl activate <name>.
- Confirm it appears as a prompt in the connected client.
Output
Return the full SKILL.md content, the path you would save it to, and a
two-line note on the scope decisions you made (what you left out and why).