skill-creator（技能创建器）

Skill Creator for z-schema

Create skills that help people and AI agents accomplish goals with z-schema — a JSON Schema validator supporting draft-04 through draft-2020-12.

Skills live in skills/<skill-name>/ and teach Claude how to perform z-schema tasks reliably: validating data, writing schemas, handling errors, using custom formats, contributing to the codebase, and more.

Communicating with the user

Match your communication style to the user's technical level. Most z-schema users are developers, but skill creation itself may be new to them. Briefly explain skill-specific concepts (frontmatter, progressive disclosure, trigger descriptions) when first introduced. Don't assume everyone knows what YAML frontmatter is — one sentence of explanation is enough.

Keep things conversational and practical. Avoid heavy jargon unless the user is clearly comfortable with it.

Creating a skill

Step 1: Capture intent

Understand what the user wants the skill to accomplish. The conversation may already contain a workflow to capture — if so, extract what you can from it before asking questions.

Determine:

1. What should this skill enable? (e.g., "validate API responses", "migrate schemas from draft-04 to draft-2020-12", "set up z-schema in a new project")
2. Who is the audience? Library consumers using z-schema as a dependency, or contributors working on z-schema's source code?
3. When should this skill trigger? What phrases or contexts should activate it?
4. What's the expected output? Code snippets, configuration, step-by-step instructions, file changes?
5. Does the skill need test cases? Skills with deterministic outputs (code generation, schema transforms, error handling patterns) benefit from tests. Skills with subjective outputs (architecture advice, code review style) usually don't.

Step 2: Interview and research

Ask about edge cases, input variety, and success criteria. Probe for specifics:

• Which JSON Schema drafts does the skill need to cover? All of them, or specific ones?
• Does it involve z-schema options, formats, remote references, error handling?
• Are there existing docs in docs/ that cover this territory? Check before reinventing.
• Is there an existing skill in skills/ that overlaps? Review it to avoid duplication and maintain consistency.

Read the relevant z-schema documentation to ground the skill in accurate, current information:

• docs/architecture.md — module structure and validation pipeline
• docs/conventions.md — code style, naming, imports
• docs/testing.md — test framework, file naming, patterns
• docs/contributing.md — PR workflow, adding features/errors/formats
• docs/usage.md — library API, validation modes, options
• docs/features.md — feature catalog with examples
• docs/options.md — full options reference

Study the existing skills/validating-json-schema/SKILL.md as a reference for tone, structure, and level of detail.

Step 3: Write the SKILL.md

See references/skill-structure.md for the full structural guide, z-schema conventions, and examples.

Core principles:

1. Start with frontmatter — name and description are required. The description is the primary trigger mechanism — make it specific and slightly "pushy" to ensure the skill activates when relevant.
2. Keep SKILL.md under 500 lines. Move detailed reference material (option tables, error code lists, schema examples) into references/ files. Point to them clearly from the main SKILL.md with guidance on when to read each one.
3. Use imperative instructions. Write "Create a validator with ZSchema.create()" not "You should create a validator...".
4. Explain the why. Instead of rigid MUST/NEVER rules, explain reasoning. Today's LLMs respond better to understanding motivation than to heavy-handed directives.
5. Include working code examples. z-schema skills are most useful when they contain copy-paste-ready TypeScript snippets. Always use ZSchema.create() (never new ZSchema()), always use the correct import style, and specify which draft the example targets if it matters.
6. Ground in z-schema reality. Every API call, option name, error code, and type name in the skill must be accurate. Cross-reference against src/index.ts exports, docs/options.md, and src/errors.ts to verify.

Step 4: Organize reference files

If the skill needs detailed supplementary material, add files under <skill-name>/references/:

skill-name/
├── SKILL.md              # Main instructions (<500 lines)
└── references/           # Detailed reference material
    ├── topic-a.md        # Loaded on demand
    └── topic-b.md        # Loaded on demand

Good candidates for reference files:

• Full option tables (see skills/validating-json-schema/references/options.md)
• Complete error code listings (see skills/validating-json-schema/references/error-codes.md)
• Draft-specific migration guides
• Large schema examples
• Step-by-step tutorials for complex workflows

Include a table of contents in reference files over 150 lines.

Step 5: Test the skill

For skills with verifiable outputs, create 2–3 realistic test prompts — things a real user would say. Share them with the user for confirmation before running them.

Test prompt examples for a z-schema skill:

• "I have user registration data coming from a form and I need to validate it has the right fields and types before saving to the database"
• "My schema uses $ref to reference a shared address definition and validation is failing with an unresolvable reference error"
• "I want to validate that dates in my API payload match ISO 8601 format"

Running tests:

Execute each test prompt yourself, following the skill's instructions to complete the task. Verify:

• Code examples compile and work with the current z-schema API
• Error codes and option names are accurate
• The skill covers the user's scenario without gaps
• Instructions are unambiguous — a developer unfamiliar with z-schema could follow them

Save test results for the user to review. Organize by test case:

<skill-name>-workspace/
└── iteration-1/
    ├── test-1/
    │   └── output/       # Generated code, schemas, etc.
    ├── test-2/
    │   └── output/
    └── notes.md          # What worked, what didn't

Step 6: Iterate

After reviewing test results with the user:

1. Generalize from feedback. Don't overfit to test cases. If a user says "the error handling example doesn't show nested errors from oneOf", the fix isn't just adding that one example — it's ensuring the error handling section covers combiner keywords comprehensively.
2. Keep it lean. Remove instructions that aren't pulling their weight. If test runs show the skill causing unnecessary steps, trim.
3. Explain the why. If you find yourself writing "ALWAYS do X", reframe: explain why X matters so the model (or human) understands the reasoning.
4. Look for repeated patterns. If every test case required the same boilerplate setup, the skill should include that setup as a template.
5. Re-verify accuracy. After changes, re-check API names, option defaults, error codes, and types against the source code.

Rerun tests after changes. Repeat until the user is satisfied or feedback is all positive.

Improving an existing skill

When the user wants to improve a skill that already exists:

1. Read the current skill thoroughly — SKILL.md and all reference files.
2. Understand the complaint — Is it inaccurate? Incomplete? Poorly triggered? Hard to follow?
3. Check against current source — z-schema's API may have changed since the skill was written. Verify all code examples, option names, error codes, and types against src/.
4. Apply the same iteration loop — make changes, test, review, repeat.

Common improvement tasks

• Skill doesn't trigger reliably: Rewrite the description in frontmatter. Make it more specific about when to activate. Include synonyms and related phrases.
• Code examples are wrong: Cross-reference against src/index.ts exports and docs/usage.md. Ensure ZSchema.create() is used, imports are correct, and the draft version matches the example.
• Missing coverage: Check what the user is asking about against what the skill covers. Add sections or reference files as needed.
• Too long: Extract detailed material into references/ files. Keep SKILL.md focused on the most common workflows.
• Outdated after z-schema changes: Read the CHANGELOG.md and recent commits. Update affected sections.

z-schema-specific guidance

Every skill for this repository should respect these constraints:

API accuracy

• Factory pattern: Always ZSchema.create(options?) — never new ZSchema(). This returns typed variants based on options (ZSchema, ZSchemaSafe, ZSchemaAsync, ZSchemaAsyncSafe).
• Default draft: draft2020-12. Always specify the draft explicitly in examples if it matters.
• Imports: Use import ZSchema from 'z-schema' (default) or import {ZSchema} from 'z-schema' (named). Use import type {...} for type-only imports.
• Error shape: ValidateError has .details (not .errors) — an array of SchemaErrorDetail with message, code, params, path, keyword, inner.

Validation modes

Skills should demonstrate the mode appropriate to the use case:

• Sync throw (default): validator.validate(data, schema) — throws ValidateError
• Safe: validator.validateSafe(data, schema) — returns {valid, err?}
• Async: requires {async: true} — for async format validators
• Async safe: {async: true, safe: true} — returns Promise<{valid, err?}>

Draft differences

If a skill involves draft-specific features, be explicit about which drafts support them:

• prefixItems — draft-2020-12 only (replaces array-form items)
• $dynamicRef/$dynamicAnchor — draft-2020-12 only
• $recursiveRef/$recursiveAnchor — draft-2019-09 only
• unevaluatedProperties/unevaluatedItems — draft-2019-09 and draft-2020-12
• if/then/else — draft-07+
• $id, const, contains — draft-06+

Code conventions for contributor-facing skills

Skills targeting z-schema contributors (not just consumers) must follow the codebase conventions:

• TypeScript with strict: true, ESM with .js import extensions in src/
• import type for type-only imports
• Tests in test/spec/ with .spec.ts / .node-spec.ts / .browser-spec.ts suffixes
• Exports through src/index.ts
• Error codes in UPPER_SNAKE_CASE in src/errors.ts

See docs/conventions.md and docs/contributing.md for the full rules.

Existing documentation

Before writing new reference material, check if docs/ already covers it:

• docs/usage.md — full library API guide
• docs/options.md — every option with description and default
• docs/features.md — feature catalog with code examples
• docs/architecture.md — module structure and validation pipeline
• docs/testing.md — test framework and patterns
• docs/contributing.md — PR workflow and code change guides

Point to existing docs rather than duplicating them. Only create skill-specific reference files when the skill needs a different angle or aggregation of information.

Skill description optimization

The description field in SKILL.md frontmatter determines whether the skill gets activated. After creating or improving a skill, review the description for triggering accuracy.

Good descriptions:

• State what the skill does AND when to use it
• Include synonyms and related phrases users might say
• Are slightly "pushy" — lean toward triggering when there's any doubt
• Mention specific z-schema concepts the skill covers

Example — weak:

description: How to validate JSON data with z-schema.

Example — strong:

description: Validates JSON data against JSON Schema using z-schema. Use when the user needs to validate JSON, define schemas, handle validation errors, use custom formats, or work with JSON Schema drafts 04 through 2020-12. Covers sync/async modes, safe error handling, schema compilation, remote references, and TypeScript types.

Tuning the description:

1. List 5–10 realistic prompts that should trigger the skill
2. List 5–10 similar prompts that should NOT trigger it
3. Check: does the description clearly include concepts from the should-trigger list while being specific enough to exclude the should-not-trigger list?
4. Revise and re-check

Reference files

For the full structural guide, conventions, and examples for writing z-schema skills:

• references/skill-structure.md — Skill anatomy, frontmatter schema, progressive disclosure, writing patterns, and complete examples

Core loop summary

1. Understand what the skill should do
2. Research the z-schema docs and source to ground the skill in accuracy
3. Draft the SKILL.md and any reference files
4. Test with realistic prompts
5. Review results with the user
6. Iterate until the skill is accurate, useful, and well-triggered
7. Verify all code examples, API names, and types against current source

Break these into tracked tasks so nothing gets skipped. Take your time writing a good draft, then look at it with fresh eyes and improve it before sharing.