Contributing¶

Workflow¶

Read the spec first: every skill or agent follows a specification.
Create the skill/agent: use the Skill Management skill. It scaffolds ASCII-kebab-case folders, writes valid frontmatter, and prevents the usual mistakes.
Adjust the spec if needed: via the Spec skill. Never edit translations directly; the canonical (EN) is the source of truth, everything else is regenerated from it.
Validate: skill-management in validation mode. Mechanical defects (frontmatter mismatch, absolute paths, missing hard-rules) can be autofixed on request.
Update the index: after spec changes, regenerate spec/README.md through the Spec skill.

Names: ASCII kebab-case.
Descriptions: concrete user triggers ("use when the user says X"), not abstract capabilities.
Agent tool access: principle of least authority. Read-only = no write tools.
No absolute paths inside skill or agent content.
Content language: skills and agents in English, user response in the user's language.

Short, imperative subject lines ("add spec skill template").
One logical change per commit. Spec changes (all languages + index update) stay together in a single commit so the canonical never lands on a branch without its translations.

Link the affected spec.
Describe which skills/agents are new or changed.
Mention what you tested (for example "loaded plugin via claude --plugin-dir ., /nolte-shared:spec shows up in the /skills dialog").

Edited a translation directly → drift. Change the canonical and regenerate.
Created placeholder subfolders → the skill-management skill forbids empty "just in case" folders.
description too vague → Claude can't route reliably. At least three concrete phrasings.
tools too generous on an agent → read-only agents get no write/execution tools, not even "for later."