Content
57%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-sectioned, code-heavy reference that is genuinely actionable for the common CLIP use cases. Its main weaknesses are marketing-flavored padding (GitHub-star metrics, repeated license info), code fragments that assume state from earlier examples, no validation guidance for batch or moderation workflows, and a references/applications.md bundle file that the body never points to despite duplicating its content.
Suggestions
Drop the "Metrics" section and the duplicate star count/license mentions (they are time-sensitive and already on the linked GitHub repo), and remove obvious comments like "# Load model" to tighten token usage.
Link references/applications.md from the body (e.g., an "## Applications" section pointing to it) and move the per-use-case recipes it duplicates — image search, content moderation, classification variants — out of SKILL.md into that file.
Add validation checkpoints: handle failed or unreadable images in the batch-processing loop, and in content moderation flag results below a confidence threshold instead of always returning the argmax category.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is mostly lean, code-first prose, but it carries unnecessary padding: a marketing "Metrics" section ("25,300+ GitHub stars", "MIT License") whose star count and license are repeated again in Resources and the frontmatter, obvious comments like "# Load model", and a duplicated label list inside the quick-start example. This matches anchor 3 (mostly efficient but could be tightened) rather than anchor 4's minor-trim level, and the time-sensitive star counts would require upkeep. | 3 / 5 |
Actionability | Installation commands and the zero-shot classification quick start are copy-paste executable, and each use case has concrete code. However, later sections (image-text similarity, content moderation, the Chroma integration) are fragments that silently depend on variables (model, image, text, image_paths, image_embeddings) defined in earlier examples, which is exactly anchor 4's "concrete code with minor gaps" rather than the fully self-contained anchor 5. | 4 / 5 |
Workflow Clarity | Each use case is presented as a short, unambiguous linear recipe, but there are no validation or verification steps anywhere: the batch-processing section does not handle failed image loads or empty batches, and the content-moderation section acts on a single argmax with no confidence threshold or uncertain-class handling. Per the rubric's cap for batch operations without feedback loops, workflow clarity cannot exceed 3. | 3 / 5 |
Progressive Disclosure | The bundle contains references/applications.md, yet the body never mentions or links it — its application recipes are instead duplicated inline (classification, image search), and Resources lists only external URLs. This matches anchor 3's "references present but not clearly signaled; content that should be separate is inline" — structure is decent, but the orphaned reference file and inline duplication prevent anchor 4. | 3 / 5 |
Total | 13 / 20 Passed |