Content
73%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.
As writing, the body is excellent — lean, dense with executable commands, and sequenced with explicit validation gates and feedback loops. As a bundle, it fails its own architecture: it is explicitly a thin index whose per-phase rules and page template are not shipped, so the progressive-disclosure structure it advertises cannot be followed.
Suggestions
Ship the referenced files — rules/analysis.md, rules/diagram.md, rules/outputs.md, and templates/page.html — or inline the phase procedures into SKILL.md if the bundle is meant to be self-contained; a thin index pointing at absent files is the skill's dominant defect.
State the consent rule once (Core rules) and reference it from the Arguments and stack sections instead of restating it four times, trimming the long --findings table cell to just the pointer into rules/diagram.md.
Add natural trigger variations ('PR overview', 'summarize this PR', 'impact analysis') to the description's trigger clause so the skill is discoverable before the user knows its name.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is dense and load-bearing — no explanation of concepts Claude already knows, compact wrong/right contrast blocks instead of prose, and exact commands rather than descriptions. It is not 5 because of repetition that could be trimmed: the GitHub-write consent rule is stated four times (Arguments prose, stack step 3, the ✗/✓ block, and Core rule 5), and the long --findings table cell re-states procedure detail that is also delegated to rules/diagram.md. | 4 / 5 |
Actionability | Concrete and executable where it speaks directly: exact script invocations ("node scripts/line-link.mjs locate pr.diff <path> '<regex>'"), gh commands ("gh pr diff <n>", "gh pr list --base <head>"), a complete flag table, a literal report template, and per-phase gates like "≤ 3 decisions, each What / Risk / Ask and ≤ 3 linked lines". It is not 5 because the procedures for phases 0–4 are delegated to rules/*.md files that are not present in the bundle, so a reader cannot execute the core analysis from what actually ships alongside this index. | 4 / 5 |
Workflow Clarity | The Workflow table sequences seven named phases (Fetch → Candidates → Verify → Decide → Diagram → Page → Post), each with an explicit gate that functions as a validation checkpoint: "Every candidate classified open / settled / wrong, with evidence", "PNGs rendered and looked at", and "PNGs uploaded, posted, read back clean" — the last two being feedback loops for error recovery. Risky operations carry their own checks (consent verified before the undeletable image upload; the no-Chrome fallback is specified), so the destructive/batch cap does not apply. Not 4 because checkpoints are not merely 'most' present but explicit at every phase. | 5 / 5 |
Progressive Disclosure | The design intent is exemplary — "This SKILL.md is a thin index", one-level-deep per-phase rule files, and an on-demand reference that self-describes as "Read on demand; nothing here is needed to run the skill" — but scored against the actual bundle, 5 of the 6 referenced paths are broken: rules/analysis.md, rules/diagram.md, rules/outputs.md (the workflow table's core targets) and templates/page.html do not exist in the bundle; only references/design-rationale.md and the two scripts are present. Navigation is clearly signaled but leads nowhere for the procedural content, which is the failure this dimension exists to catch. Not 3 because the missing files are the skill's substance, not an organization nicety; not 1 because the in-body structure and the one real reference are well organized and only one level deep. | 2 / 5 |
Total | 15 / 20 Passed |