Write reusable guidance around decisions, evidence and useful outputs. Explain when a method helps, how to recognize failure and when a simpler approach is better. Avoid mandatory phases, handoff schemas, arbitrary scores and prose that merely repeats another skill. Link the owning skill instead.

Every skill has SKILL.md frontmatter with a equal to its directory and a trigger-focused starting with . Its agents/openai.yaml contains display_name, short_description and a default_prompt mentioning the matching . These metadata files are discovery interfaces, not a workflow engine. Use concise role definitions; inherit the host's model choice.

Put detailed figure recipes, source provenance and examples in references or examples directories within the owning skill. Keep the entry point short and load only the references relevant to the task. Preserve third-party licensing; never execute downloaded source snapshots as an installation step.

Small scripts are appropriate when they perform concrete work such as plotting, backup or rendering. Keep them optional, declare dependencies, and verify the behavior they actually promise. Do not create a framework to enforce prompts.

Update all affected entry points, role skill lists, metadata and English/Chinese documentation. Keep each language natural; heading or code-block counts need not match. Check that commands, capabilities and links agree. Use maintenance for mechanical checks.