Agent Skills
› krafton-ai/KIRA
› skill-creator
skill-creator
GitHub用于创建或优化可复用的 Agent Skill 包。指导定义触发条件、确定范围、编写工作流与规则,确保技能简洁高效,避免冗长说明,提升模型加载性能与执行可靠性。
Trigger Scenarios
用户要求制作新技能
更新现有技能
需要可复用的代理工作流指令
Install
npx skills add krafton-ai/KIRA --skill skill-creator -g -y
SKILL.md
Frontmatter
{
"name": "skill-creator",
"description": "Create or refine KiraClaw\/Codex skills as concise SKILL.md packages. Trigger on: \"make a skill\", \"create a skill\", \"update this skill\", \"seed skill\", \"default skill\", or when the user wants reusable agent workflow instructions instead of one-off prompt text."
}
Skill Creator
Use this skill when the task is to create or refine a reusable skill package.
Goals
- keep the skill short enough to load cheaply
- encode workflow, not generic model intelligence
- use references or scripts only when they clearly reduce risk or repetition
- fit KiraClaw's seeded-skill layout under
Filesystem Base Dir/skills
Workflow
- Define the trigger clearly:
- what user requests should activate this skill?
- what should not activate it?
- Decide the scope:
- one focused workflow per skill
- avoid combining unrelated jobs into one package
- Write the frontmatter carefully:
namedescription- include trigger examples in the description
- Write the body as operational guidance:
- workflow steps
- failure modes
- decision rules
- exit criteria
- Add references or scripts only if they reduce repeated work or improve reliability.
Rules
- Prefer one strong skill over one huge omnibus skill.
- Do not explain concepts the base model already knows unless the local workflow is special.
- Do not turn the skill into a README or changelog.
- If the skill depends on local tools, directories, or conventions, say so directly.
- Keep examples short and realistic.
Good Skill Shape
Most good skills have:
- a narrow trigger
- a short workflow
- a few rules that prevent common mistakes
- one or two high-value examples
When to Split
Split the skill if:
- it has multiple unrelated trigger families
- it needs different rules for different domains
- the body is becoming reference material instead of workflow guidance
Seed Skill Guidance
For seeded default skills:
- optimize for broad usefulness
- avoid org-private assumptions
- keep the body compact so new workspaces do not inherit unnecessary prompt weight
Validation
Before finishing a new skill, check:
- is the trigger description specific enough?
- does the body contain real workflow guidance?
- can at least one realistic task use it without extra explanation?
Version History
- 652dacb Current 2026-08-20 12:24


