author-skills

Front-load the skill's leading word and list one trigger per genuinely distinct branch, including quoted trigger phrases. Test whether a smaller model would know when to load the skill from the description alone; collapse synonym triggers as duplication.

Read each sentence in isolation and ask whether the model behaves differently because that line exists. If not, delete the whole sentence rather than trimming words. Weak leading words like 'be thorough' are no-ops; replace them with stronger concrete terms.

Use the scrappy lint pass for one-line edits or quick triage of an already-shipped skill. Use the full rigor path, including the complete failure-mode QA, for any new or forked skill entering the library, since it becomes permanent infrastructure.

This rubric deliberately keeps prohibitions as load-bearing guardrails for smaller models, rejecting the source rule that bans negation. Keep Don't/Do tables where every prohibition is paired with its positive alternative; prefer positive phrasing only inside prose steps.

Restated values drift across copies, which is the failure the library exists to prevent. Replace any restated taste, evidence, or bias content with a pointer such as 'load design-taste first' so a single source of truth owns each value.

Keep the body between 120 and 350 lines with at least one complete worked example. If content exceeds the limit, move reference material into a references file behind a context pointer rather than deleting live, unique content.