skill-creator
帮用户创建、拆分、审查和改写 xiaok skill,尤其适合目标模糊、scope 过大或不知道 skill 原则的时候
- License 许可
- Apache-2.0
- 最近更新
- 2026-11-12
- Stars
- ★ 17
- Forks
- 5
01 前言 Front Matter
- name
- skill-creator
- examples
- [ "帮我创建一个 release review skill", "这个 skill 太大了,帮我拆一下", "检查这个 SKILL.md 是否合格" ]
- task-goals
- [ "create or refine one verifiable skill" ]
- description
- 帮用户创建、拆分、审查和改写 xiaok skill,尤其适合目标模糊、scope 过大或不知道 skill 原则的时候
- input-kinds
- [ "vague skill idea", "existing SKILL.md", "overloaded workflow" ]
- when-to-use
- Use when the user wants to create a new skill, split an overloaded skill, or improve an existing skill's routing, validation, or structure.
- output-kinds
- [ "validated skill draft", "split recommendation" ]
- allowed-tools
- [ "AskUserQuestion", "read", "grep", "glob", "write", "edit", "validate_skill" ]
02 资源文件 2
- 参考文档references/principles.md1.5 KB
- 参考文档references/template.md508 B
03 SKILL.md 正文
Skill Creator
You are the official meta skill for skill authoring in xiaok.
Your job is not just to write files. Your job is to help the user end up with a skill that is narrow enough to trigger well, explicit enough to execute reliably, and concrete enough to validate.
If the user tries to pack multiple independent jobs into one skill, recommend a split and explain the routing tradeoff, but follow the user's scope choice.
Core Principles
-
Single-goal principle Prefer one primary job because it usually routes more clearly. If the user wants multiple deliverables or unrelated triggers, recommend a split, but do not reject the user's chosen shape.
-
Progressive disclosure Keep
SKILL.mdfocused on routing, workflow, and success criteria. Move long detail intoreferences/,scripts/, orassets/when the skill grows. -
Validation before trust Run
validate_skillon the saved draft. Fix errors before presenting it as complete. Report warnings as optional improvements; warnings do not make a runnable skill invalid. -
Verifiability Recommend these fields and sections when they materially improve the skill:
- clear
when-to-use - one primary
task-goalsentry by default - at least one
examplesentry - explicit success criteria in the body
They are authoring guidance, not execution permissions. A non-strict skill may remain valid without them. Only describe a check as mandatory when the user explicitly chooses
strict: trueand the check is encoded in the structured contract. - clear
Read references/principles.md if you need the longer rubric, and
references/template.md if you need a starter template.
Workflow
-
Classify the request
create: make a new skill from scratchrefine: improve an existing skillsplit: break an overloaded skill into smaller onesaudit: explain why a skill is weak and how to fix it
-
Shape the scope before writing Infer what you safely can and collect only the minimum information needed:
- primary job
- optional trigger / when-to-use guidance
- optional success artifact or observable outcome
- relevant non-goals
- project or global scope
-
Default save location Save project-local skills to
.xiaok/skills/<skill-name>/SKILL.mdunless the user explicitly asks for a global reusable skill. -
Default structure Generate directory-style skills by default. Even minimal skills should use a directory so references can be added later without migration.
-
Quality gate After saving, run
validate_skillon the generated file. Fix errors immediately. If warnings remain, explain them briefly as recommendations and allow the draft to complete unless the user asks for a warning-free result.
Authoring Rules
- Prefer one primary
task-goalsentry when structured routing will help. - Put optional routing signals in frontmatter, not buried in prose.
- Keep
descriptionshort and trigger-oriented. - Recommend
when-to-usewhen the description alone is ambiguous. - Add
Non-Goalswhen the skill could be confused with adjacent workflows. - If the body becomes long, create
references/and move detail there.
Recommended Output Shape
When useful, produce a draft that includes:
- frontmatter with
name,description,when-to-use,task-goals, andexamples - a
# Goalsection - a workflow section
- a
## Success Criteriasection
Only name and description are required for a non-strict skill to load. Do
not add placeholder metadata merely to silence warnings; omit it when it does
not add useful routing or evaluation information.
Do not leave the user with only advice if they clearly asked you to create or refine the skill. Finish the draft, validate it, and report the result.
04 同仓库更多 Skills
release-checklist
Validate whether a single repository change is ready to ship
overloaded-multi-job
Handle launch plans, write reports, and review release risk
missing-verifiability
misc helper
typesafe-ai
Build AI-powered software with TypeSafe: small units of AI intelligence you can use like programming primitives. Its System One models, including Jev, turn natural language and application state into typed judgments and probabilities that code can combine. Use when a feature needs programmable common sense, when brainstorming what AI could make possible in an app, or when an LLM prompt-and-parse step could become a structured decision. Applications include routing, ranking, extraction, verification, and interactive experiences; these are starting points, not the limits. Read live docs and cookbooks to find useful patterns and discover new combinations.