hermes-agent-skill-authoring

Write the SKILL.md with write_file to skills/<category>/<name>/ or optional-skills/<category>/<name>/, then git add and commit. Do not use skill_manage(action='create'), which only targets the user-local ~/.hermes/skills/ tree.

The validator requires name and description, but repo review also expects version, author, license, platforms, and metadata.hermes with tags and related_skills. The file must start with --- at byte zero and close the frontmatter before a non-empty body.

The repo hardline is 60 characters, one sentence, ending with a period, even though the validator allows 1024. The system prompt index truncates at 57 characters, so the trigger must be self-contained in that window.

Bundled skills/ is for daily-driver behavior useful across many user types with a 5+ sessions per month bar. Niche, vertical-specific, or heavy skills belong in optional-skills/; when in doubt, choose optional since promoting later is easier than demoting.

The skill loader is initialized at session start and cached, so skill_view and skills_list will not see a newly committed skill until a new session begins. This is expected behavior, not a bug.

Audit what the skill's prose and scripts actually invoke rather than copying a sibling. Cross-platform Hermes tools and stdlib Python allow all three platforms, while bash pipelines limit to linux and macos, and osascript or apt narrow to a single OS.