transformer
This skill contains shell command directives (!`command`) that may execute system commands. Review carefully before installing.
input = $ARGUMENTS
Rewrite the target skill so Opus 4.8 executes it with better judgment. Resolve $input to a path; if no file there, try ~/.claude/skills/$input/SKILL.md then ./.claude/skills/$input/SKILL.md — if none resolve, ask. Read it completely. Write the rewrite in place — the diff is the review surface.
The one test
For every line: what wrong behavior happens if this is gone? Wrong output, skipped coordination, lost trigger, silent regression → keep. The model arrives here unaided → strip. Length is a side effect of this test, not a target — "nothing to cut, much to add" is a valid verdict, and a rewrite that removes more than it adds is not automatically better.
The 4.8 strip/keep lists, the base tendencies worth countering, and the effort-vs-scaffolding rule are in the inlined reference below — apply them.
Both-directions guardrail
Under-specification is the twin failure of over-scaffolding. If the target is principle-only with no anchoring, ADD: one example showing the core judgment call; 2–3 domain facts the model can't infer, phrased positively; explicit invariants, XML-tagged when load-bearing.
Preserve verbatim: frontmatter description trigger phrases + Do NOT trigger disambiguation, allowed-tools, output contracts/templates, approval gates, and !`cat path`` inline-injection directives. Breaking any of these breaks the skill silently.
Worked transformation
Before — over-scaffolded: