Content
46%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.
The body has a genuinely good progressive-disclosure architecture with six real, well-signaled reference files and substantive API coverage, but it is undermined by heavy auto-generated boilerplate, duplicated trigger sections, code examples with undefined variables, and missing validation/error-recovery checkpoints. The first quarter of the file is padding that adds no actionable value.
Suggestions
Delete the boilerplate sections: the duplicated description in 'When to Use' and 'Key Features', the fake 'Example Usage' directory listing, the 'Implementation Details' section that points to '## Overview above', and the trailing K-Dense Web advertisement — keep a single concise 'When to Use' section.
Make the four use-case examples self-contained and executable: define the 'schema' variable in Use Case 1, initialize 'data_files'/'tissues'/'conditions' in Use Case 2, replace the undefined 'train_model' call with a real estimator (e.g., sklearn) in Use Case 3, and label Use Case 4 explicitly as a Nextflow template.
Add explicit validation checkpoints with error-recovery loops to the curation and data-registration workflows (e.g., 'run curator.validate(); if it fails, standardize with .cat.standardize() and re-validate before save_artifact()').
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The first ~50 lines are auto-generated boilerplate: the frontmatter description is repeated verbatim (lowercased) in 'When to Use' and again word-for-word in 'Key Features' ('Scope-focused workflow aligned to: This skill is applicable when using LaminDB...'), followed by filler like the fake directory listing in 'Example Usage', an 'Implementation Details' section that just says 'See ## Overview above for related details', and an unrelated K-Dense Web advertisement at the end. There are also two overlapping 'When to Use' sections. Not 1 because the core capability sections and code examples do carry real, non-redundant information; firmly at 2 because several padded sections teach nothing Claude doesn't already know. | 2 / 5 |
Actionability | Concrete guidance exists — install commands ('uv pip install lamindb', 'lamin login', 'lamin init --storage ...'), API calls ('ln.track()', 'bt.CellType.import_source()', 'curator.validate()', 'ln.Artifact.filter(key__startswith="scrna/", ...)') — but the use-case examples are not executable as written: undefined 'schema' in Use Case 1, undefined 'data_files', 'tissues', 'conditions' in Use Case 2, undefined 'train_model' in Use Case 3, and Nextflow template pseudocode ('${batch_id}') in Use Case 4. Not 4 because multiple examples fail as copy-paste code due to undefined variables; not 2 because there is substantial specific, correct API usage beyond high-level hints. | 3 / 5 |
Workflow Clarity | The 'Getting Started Checklist' provides a reasonable six-step learning sequence and 'Key Principles' gives ordering guidance ('Validate early', 'Query metadata first'), and curator.validate() appears in Use Case 1. But no workflow includes error-recovery checkpoints — there is no 'if validation fails, fix and re-run' loop, and the curation section lists validation steps without any failure handling — so checkpoints are implicit at best. Not 4 because validation/feedback loops for the curation and pipeline workflows are absent; not 2 because sequences are present and reasonably defined. | 3 / 5 |
Progressive Disclosure | Structure is good: six one-level-deep reference files (core-concepts.md, data-management.md, annotation-validation.md, ontologies.md, integrations.md, setup-deployment.md) all exist and each capability section clearly signals its reference ('Reference: references/... - Read this document for...'), plus a consolidated 'Reference Files' index. Not 5 because SKILL.md itself runs ~430 lines with each capability summarized inline before pointing to its reference, and the boilerplate front sections ('Example Usage', 'Implementation Details', duplicated 'When to Use') add navigational noise; not 3 because references are clearly signaled, shallow, and well-indexed. | 4 / 5 |
Total | 12 / 20 Passed |