Nima Skill Creator

Nima Skill Creator

Design skills as portable workflow packages, not product-specific prompt files.

Operating principles

1. Start from concrete requests, expected outputs, and failure cases.
2. Use the Agent Skills open specification as the core contract.
3. Keep product-specific behavior in optional adapters or clearly marked extensions.
4. Put deterministic operations in scripts/, detailed knowledge in references/, and reusable output material in assets/.
5. Keep SKILL.md concise and link directly to supporting files.
6. Validate, test, and package before claiming the skill is complete.

Workflow

1. Audit or discover

For an existing skill, read every file that can affect execution before proposing changes. For a new skill, collect only the missing information:

• 2–4 representative user requests
• required inputs and outputs
• trigger wording and non-trigger examples
• environment, tools, permissions, and safety boundaries
• reusable scripts, references, templates, or examples

Do not repeat questions the user has already answered. When the request is sufficiently concrete, summarize assumptions and proceed.

Use interaction-guide.md for concise Chinese discovery prompts.

2. Choose a design pattern

Select the smallest useful combination from design-patterns.md:

• tool-wrapper for domain or tool guidance
• generator for stable output shapes
• reviewer for criteria-driven evaluation
• inversion for requirement discovery before execution
• pipeline for ordered stages and gates

Use multiple patterns only when each controls a distinct failure mode.

3. Plan portable resources

Use best-practices.md and compatibility.md.

• Core: SKILL.md, plus optional scripts/, references/, and assets/
• OpenAI/Codex presentation metadata: optional agents/openai.yaml
• Product-specific frontmatter: add only when requested and document the portability impact
• Repository files such as README.md, tests, CI, and contribution docs must remain outside the distributable skill package

4. Implement

Create a portable skill without product adapters:

python3 scripts/init_skill.py my-skill \
  --path /path/to/skills \
  --description "What the skill does and when to use it" \
  --resources scripts,references,assets

Add the optional OpenAI adapter only when needed:

python3 scripts/init_skill.py my-skill \
  --path /path/to/skills \
  --adapter openai \
  --interface display_name="My Skill" \
  --interface short_description="Create and improve My Skill workflows"

When updating an existing skill, make the minimum coherent set of changes and preserve product extensions that are intentional.

5. Validate and test

python3 scripts/validate_skill.py /path/to/skill

Use --strict when the target accepts only open-standard frontmatter fields. Run bundled tests or scripts and report failures honestly.

6. Package or install

Create a clean archive that excludes repository-only files:

python3 scripts/package_skill.py /path/to/skill ./dist

Install a directory or archive:

python3 scripts/install_skill.py ./dist/my-skill.skill.zip --target claude-code
python3 scripts/install_skill.py ./my-skill --target opencode --scope project --project-root /path/to/repo

Supported installer targets are documented in compatibility.md.

Review gate

Before completion, verify:

• directory name matches frontmatter name
• description states both capability and trigger scenarios
• optional standard fields are valid and product extensions are intentional
• linked files exist and do not escape the skill directory
• scripts parse and were executed when execution is material to correctness
• no secrets, credentials, symlinks, generated caches, or repository-only files enter the package
• installation instructions match the selected product and scope

Response format

Report:

1. task and compatibility summary
2. findings ordered by severity
3. files created or changed
4. commands and test results
5. remaining product-specific limitations

Use output-patterns.md for compact report shapes.