Content
88%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 dense, well-sequenced orchestration skill: every step carries concrete commands, fallbacks, and verbatim prompts, with error-recovery loops at each fragile point and reference content correctly pushed to just-in-time lens files. The main costs are a duplicated subagent-constraint quote between steps 5 and 6, labyrinthine nested sentences in steps 2 and 5, and no explicit listing of the shipped lens reference files.
Suggestions
In step 6, drop the full re-quote of the subagent constraint after "the same constraint as independent lenses:" — the reference back is sufficient and saves ~50 tokens.
Split the mega-sentences in steps 2 and 5 (e.g. the staging and empty-input rules) into short bullet sub-steps; the parenthetical nesting obscures otherwise clear rules.
Add a one-line list of the shipped lens reference files (or a note that customize.toml's lens entries name their instruction files) so the bundle's structure is discoverable from SKILL.md itself.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is lean and assumes Claude's competence ("Report what is real — never pad to look thorough"), with no explanation of known concepts. However, step 6 re-quotes the full subagent constraint that step 5 already establishes, and steps 2 and 5 are heavily nested parenthetical sentences that could be split, keeping it below anchor 5. | 4 / 5 |
Actionability | Fully executable guidance: the exact resolve_customization.py invocation, a copy-paste-ready fallback JSON array for empty/undecodable input, verbatim subagent prompt text, and precise decision rules (branch = diff against merge base, untracked files included, claims staged separately). This matches the instruction-only analogue of copy-paste-ready code. | 5 / 5 |
Workflow Clarity | Seven clearly sequenced steps with explicit error-recovery feedback loops at every fragile point: script-not-found → offer setup; other failure → read customize.toml directly; empty input → exact JSON response and stop; dependent lenses wait on `after` targets with defined behavior when the target produced nothing. Operations are read-only staging, so no destructive-validation cap applies. | 5 / 5 |
Progressive Disclosure | SKILL.md is an orchestrating overview with lens details correctly split into references/ and explicit just-in-time loading ("load only what runs"), over a flat bundle. Minor gaps: the body never enumerates the shipped lens reference files for discoverability (naming one only as a path-convention example), and lens-prose/lens-structure chain a second hop to editorial-common.md / structure-models.md — a shared-common pattern rather than buried content, so this stays at anchor 4 rather than 3 or 5. | 4 / 5 |
Total | 18 / 20 Passed |