WritingGreatSkills

Run the eight-step review pass: state the skill's root idea in one leading word, apply the no-op test line by line, collapse duplicated meaning, rewrite negations as positive targets, and verify each step ends on a checkable criterion. Report every cut tagged with the failure mode it fixed.

The six named modes are premature completion, duplication, sediment, sprawl, no-op lines, and negation. Each has a specific fix, such as sharpening completion criteria, enforcing a single source of truth, or prompting the positive behavior instead of prohibiting.

Move content down a tier when only some invocation branches need it, keeping inline what every branch requires. Phrase the pointer as an instruction, such as telling the agent to load the reference file before a specific judgment, so it fires reliably.

This is premature completion caused by a vague step-ending criterion. Sharpen the criterion so it is checkable and exhaustive, and if that fails, split post-completion work into a separate step so it stays out of view until the first step truly closes.

A leading word is a compact concept from the model's pretraining, like tight or red, that anchors a region of behavior in few tokens. It replaces restated phrases, cutting token count while sharpening the behavioral hook.