---

name: skill-creator-plus description: Create or review a reusable skill (SKILL.md) that packages a workflow, and decide whether the request should be a skill instead of a prompt, instruction, agent, or hook. Use when creating a new skill, extracting a workflow from a conversation, updating an existing skill, reviewing SKILL.md quality, or fixing weak skill triggering. Triggers on "create skill", "/create-skill", "new skill", "review skill", "fix skill trigger", "SKILL.md", "スキル作成". argument-hint: "作りたい skill の目的、trigger、入れたい resources" user-invocable: true license: Apache-2.0 metadata: author: yamapan

Skill Creator+

Design and review reusable skills that trigger reliably and stay lean.

Decision Flow

Start by deciding whether the user really needs a skill.

Need	Use
Reusable multi-step workflow with bundled scripts, references, or templates	Skill
Single focused slash task with parameterized input (Local harness only; deprecated for Agent Host)	Prompt
Always-on or file-scoped guidance	Instruction
Persona, tool restrictions, delegation, or handoffs	Custom Agent
Deterministic enforcement or lifecycle automation	Hook

If the answer is not Skill, stop and create the right primitive instead.

上表は「skill にすべきか」の即時ゲート。

→ references/customization-primitives.md for the full selection guide

When to Use

• Create skill, /create-skill, new skill, review skill, fix skill trigger, SKILL.md, workflow, スキル作成
• Creating a new skill from scratch
• Extracting a repeated workflow from a conversation, incident, or checklist
• Updating or refactoring an existing skill
• Reviewing existing SKILL.md files
• Deciding whether a customization should be a skill before authoring it

Start Here

1. Extract the workflow you want to package.
2. Clarify only the missing dimensions: target outcome, personal vs workspace scope, checklist vs full workflow.
3. Draft the smallest useful SKILL.md, then move detail into references.

License And Validation Gate

Classify the target before creating or changing it: self-authored, third-party, or derivative. A self-authored target uses independent content; copying this Skill's prose, code, templates, or notices makes it a derivative and requires preserved upstream evidence.

For a new self-authored Skill, propose CC BY-NC-SA 4.0 as a candidate, show the exact display attribution, and get confirmation. Do not silently copy this Skill's Apache-2.0 license into the target. Preserve an existing or upstream license when updating or importing.

Use the target license profiles and validation lanes. Baseline review is always required; Python helpers and packaging are conditional.

Gotchas

• init_skill.py --path takes the existing parent directory; the target becomes <path>/<skill_name>. Passing the skill directory itself nests a duplicate.
• Three artifacts are generated: SKILL.md, LICENSE.txt, skill-license.json.
• quick_validate.py compares LICENSE.txt content exactly against the rendered template and requires metadata.author to equal authorAttribution.value character for character. Never hand-edit LICENSE.txt; regenerate it with the tool and copy the result over.
• A passing gate proves consistency, not correctness. The manifest hashes whatever the template rendered, so a broken template yields a broken notice that still validates. Read the rendered LICENSE.txt once after scaffolding.
• Classification turns on whether the source is independently authored, not on the act of copying. Reusing your own material stays self-authored; PROVENANCE.md and upstreamAttributions are required only for third-party and derivative.
• A Skill with no skill-license.json is treated as legacy and skips the content comparison, so repairing a shared template does not invalidate older Skills. Skills that do have a manifest fail with profile snapshot disagrees after a template change: re-render LICENSE.txt and the profile snapshot (self-authored) or refresh only the snapshot (derivative) with the toolchain.
• When importing a Skill into another repository, regenerate LICENSE.txt and skill-license.json with that repository's toolchain. The manifest pins the template hash, so a copied pair fails validation there.
• Before uploading to Microsoft 365 Agent Builder (Preview, Frontier-only), desk-check against the documented limits: 8 skills per agent, zip 50 MB, SKILL.md under 20,000 characters, directory depth 3, 350 files, scripts only .py .js .mjs .cjs .ts .mts .sh .bash (no .ps1) with no network or installs in the sandbox. package_skill.py nests files under <skill_name>/ while the documented example has SKILL.md at the zip root; import of either layout is untested, so do not promise it. Source: learn.microsoft.com/microsoft-365/copilot/extensibility/declarative-agent-skills

Core Principles

Principle	Description
Concise is Key	Context window is shared. Only add what Claude doesn't already know.
Discovery First	The description is the routing surface. Triggers must be explicit.
Degrees of Freedom	Match specificity to task fragility (high/medium/low freedom)
Progressive Disclosure	Split into 3 levels: Metadata → Body → References
Integrate Before Add	Update, merge, or replace existing guidance before appending more.
Right Primitive	A good skill is not a fallback for prompt/agent/instruction design.
Scope Before File	Decide workspace vs profile before creating anything.
Self-Contained	The skill must carry its own knowledge. Bundle into references/scripts; do not just link to workspace files (instructions, memory, ledgers) that die when copied elsewhere. Abstract env-specific values (paths, names) into args/config.

Default assumption: Claude is already very smart. Challenge each piece: "Does this justify its token cost?"

Before adding a new section, ask whether it can replace an existing rule, move to references/, or be dropped as session-specific.

Skill Structure

Keep routing and decisions in SKILL.md; put deterministic helpers in scripts/, on-demand detail in references/, and reusable outputs in assets/. References load on demand and are often never read, so any constraint that gates an irreversible action — allowlists, forbidden values, safety limits — stays inline in SKILL.md even when it reads like detail, with only its rationale in references/. See Skill Structure, including excluded files.

Creation Process

Use Creation Process: choose primitive and scope, extract the reusable workflow, plan resources, implement, validate, and test real trigger prompts.

Refactor Order

When improving an existing skill, use this order:

1. Delete unnecessary or self-evident text before shortening, only if normal/exception/stop decisions stay unchanged. Preserve evidence, nontrivial procedures, scope, required examples, metadata, approvals, unverified requirements and failure/recovery conditions; age alone does not obsolete history, and unsaved notes stay protected.
2. Merge duplicate rules
3. Move long detail to references/
4. Add genuinely missing guidance last

Frontmatter and Triggering

Use the smallest viable frontmatter: name, description, and only behavior-changing optional fields. name must match the folder and description must include trigger conditions.

Frontmatter details and silent failures: references/common-pitfalls.md

SKILL.md Guidelines

Keep SKILL.md lean: <150 lines is good, >300 lines must split to references. Start When to Use with user phrases, keep essential workflow only, and push long examples / schemas / recipes to references.

Detailed review criteria: references/skill-review-checklist.md

Iteration Loop

1. Draft the skill and save it.
2. Identify the weakest or most ambiguous parts.
3. Tighten those parts, then summarize what the skill produces and example prompts to try.

Key References

Use primitive choice, structure, creation, review checklist, bloat review, and common pitfalls. The structure gallery, workflows, and output patterns are optional examples.

Done Criteria

• Request is confirmed to be a skill, not another primitive
• Scope is decided before file creation
• SKILL.md created and under 150 lines
• Frontmatter has name + description with trigger conditions
• Manually used skills have argument-hint
• user-invocable is set intentionally
• Optional fields are added only when they change behavior
• Details moved to references/ (Progressive Disclosure)
• Self-Contained: no hard reference to other skills / workspace files; env-specific values (paths, customer names, tenant IDs) abstracted
• Review checklist passed