Content
60%Weight 40%Scale 1-5Reviews the quality of instructions and guidance provided to agents. Good implementation is clear, handles edge cases, and produces reliable results.
Operationally rich with real commands, templates, and a well-sequenced gated workflow for a destructive in-place operation, but held back by duplicated flow-resolution logic and a bundle whose referenced pattern and flow files are entirely missing, leaving the actual humanization rules inaccessible.
Suggestions
Ship the referenced files (references/ai-patterns.md, references/human-patterns.md, references/flows/*.md) or inline their essential rules, since the two-pass strip/inject model is not executable without them.
Deduplicate the flow-resolution and user-gate logic: state it once in Phase 0 and have the Prompt contract section reference it instead of restating it.
Merge the two overlapping intro sections ('# Humanize text' and '# Text Humanizer') into one, and add a post-edit verification step (e.g., reviewing the diff for preserved meaning) to close the workflow feedback loop.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The flow-resolution and user-gate logic is stated twice (Prompt contract items 1/3 vs. "Argument parsing" step 4 and "Detection signals"), there are two overlapping intro sections ("# Humanize text" vs. "# Text Humanizer"), and the "Complete native workflow" paragraph is dense enough to need tightening. Everything is operational rather than padded, so it sits above 2 but below the "minor instances" bar for 4. | 3 / 5 |
Actionability | Concrete and mostly executable: the exact command `git status --porcelain -- '<resolved path>'`, copy-paste PLAN/spawn-brief/report templates, a full flow table, and eight example invocations. The gap keeping it from 5 is that the core editing rules live in `@reference/ai-patterns.md` and `@reference/human-patterns.md`, which are not present in the bundle, and "Codex collaboration" assumes unspecified tooling. | 4 / 5 |
Workflow Clarity | Clear Phase 0 -> 0.5 -> 1 -> Output sequence with an explicit destructive-operation precondition gate ("Empty -> proceed", "Unanswered -> only the clean ones; none clean -> stop, no edits") and an error-handling table ("never count unverified work done"), so the destructive/batch validation cap does not apply. Missing a post-edit verification loop, so it does not reach 5. | 4 / 5 |
Progressive Disclosure | The body references `@reference/ai-patterns.md`, `@reference/human-patterns.md`, and five `@reference/flows/*.md` files, but the bundle contains no references/ directory at all -- every referenced path is dangling, so the one-level-deep design leads nowhere. SKILL.md itself is well-sectioned (not a monolithic wall), which keeps it above 1. | 2 / 5 |
Total | 13 / 20 Passed |