Content
81%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 strong, highly actionable body: every stage is executable, the workflow is clearly sequenced with an explicit visual-diff validation gate, and references are well-signaled and one level deep. The main costs are token duplication between the Principles/When-Not-to-Use sections and the Anti-Patterns section, and a body that carries detail the existing references could absorb.
Suggestions
Consolidate the duplicated storage rules: fold the "NEVER use /tmp" and "NEVER save artifacts to docs/" anti-patterns into the Artifact Storage Convention, and drop the "NEVER pre-load all three reference files" anti-pattern since the Reference Loading Schedule table already mandates the same behavior — this trims roughly 40 lines of repetition.
Merge the obfuscated-class-names item in "When Not to Use" with the "NEVER port obfuscated class names" anti-pattern, keeping only one of the two statements of the rule.
Move the large Stage 1 inline JS payloads (1b computed-styles and the fallback extractor) into references/extraction.md, leaving one representative payload inline and pointing to the reference for the full toolkit, which the body already promises ("full JS extraction toolkit").
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | Mostly efficient — no filler explaining concepts Claude already knows, and the code blocks earn their tokens — but several rules are stated twice: the artifact-location rule appears in the Artifact Storage Convention plus the "/tmp" and "docs/" anti-patterns, the reference-loading schedule appears as both a table and an anti-pattern, and the obfuscated-class "When Not to Use" item duplicates the "NEVER port obfuscated class names" anti-pattern. This duplication is more than the "minor instances that could be trimmed" of anchor 4, so it sits at 3. | 3 / 5 |
Actionability | Fully executable throughout: copy-paste agent-browser commands, complete JS extraction payloads (with fallbacks for CSS-var-less and SPA sites), curl/grep pipelines, a complete Tailwind v4 `index.css` block, and before/after component code covering the common cases. | 5 / 5 |
Workflow Clarity | Four explicitly sequenced stages (Extract → Document → Apply → Verify), a method-selection decision tree, a required-output-files table, and an explicit validation gate — "the workflow is not complete until a diff screenshot confirms the local build matches the reference site" — with symptom→fix error recovery well-signaled at the verify stage. | 5 / 5 |
Progressive Disclosure | Three real, one-level-deep reference files are clearly signaled via both a References table (with when-to-use) and a per-stage loading schedule, which is excellent. However, the 536-line body retains large inline JS payloads and eleven anti-patterns whose extended detail could partly live in the already-existing reference files — good structure with more than minor organization gaps, so anchor 4 rather than 5. | 4 / 5 |
Total | 17 / 20 Passed |