Content
77%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.
A well-architected, terse slash-command index with an explicit phased workflow, strong validation checkpoints, and genuine load-on-demand discipline. Its main weaknesses are minor redundancy between the Workflow and Required Reading tables and — more seriously — that nearly all referenced rule files are missing from the bundle, leaving the skill's core mechanics unreachable.
Suggestions
Include the seven missing rules/*.md files (hard-rules, measurement, classification, invocation-review, audit-mode, trim-mode, extract-mode) in the bundle, or inline the critical gates (e.g. the preservation invariant) so the workflow is executable without them.
Merge the 'Workflow' and 'Required Reading by Phase' tables into a single table (Phase | Name | Rule file | Gate) and drop the repeated 'load on demand' admonitions to remove the duplication.
Reduce reliance on cross-skill paths (../create-skill/rules/invocation-control.md, ../docs/SKILL.md) by summarizing the invocation-conversion criteria inline, since those files are outside this skill's control and may not exist in the user's installation.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is lean and assumes Claude's competence (no basic-concept explanations, terse tables and one-liners), but the Workflow table and the 'Required Reading by Phase' table duplicate the same phase-to-file mapping, and 'load on demand — do not preload' is stated three times (blockquote, workflow section, required-reading header). Not 5: these redundant passages could be trimmed; not 3: everything else earns its tokens. | 4 / 5 |
Actionability | Concrete, executable directives throughout: exact '$ARGUMENTS' parsing rules, a mode table with explicit trigger keywords, a required one-line output format ('Mode: audit / Target: /abs/path/to/CLAUDE.md (43,012 chars / ~10,750 tokens)'), per-phase gates, and a Definition of Done checklist. Not 5: the actual report format and trim/extract mechanics are deferred to rules/ files that are not present in the bundle, so an agent could not fully execute from what is written. | 4 / 5 |
Workflow Clarity | The five phases (0-4) are explicitly sequenced with named gates, a preflight refusal rule ('< 10k chars'), user approval before any diff is applied, a Phase 4 verification gate ('Before/after metrics shown; no content lost in trim or extract'), an abort-and-ask fallback ('Never delete silently'), and a completion checklist — explicit validation steps plus feedback loops, matching the top anchor. The destructive/batch cap does not apply since validation is present. | 5 / 5 |
Progressive Disclosure | The design is exemplary — a self-declared 'thin index' with one-level-deep, clearly signaled, per-phase load-on-demand references — but scored against the actual bundle structure, seven of the eight referenced local files (all rules/*.md: hard-rules, measurement, classification, invocation-review, audit-mode, trim-mode, extract-mode) do not exist in the bundle; only references/bloat-patterns.md is present. Not 4: the referenced navigation does not resolve in the provided bundle; not 2: the structure that is present (including bloat-patterns.md, itself one level deep) is well organized. | 3 / 5 |
Total | 16 / 20 Passed |