Content
80%Reviews the quality of instructions and guidance provided to agents. Good implementation is clear, handles edge cases, and produces reliable results.
A concrete, well-sequenced, SQLMesh-specific walkthrough with executable commands and examples throughout and no generic-concept padding. Its main gaps are the absence of explicit error-recovery feedback loops for batch/destructive operations and a monolithic structure that would benefit from bundled reference files.
Suggestions
Add explicit error-recovery feedback loops for the batch/destructive operations: after `sqlmesh plan`, state what to do when a breaking change appears (accept and backfill, or revise the model and re-plan); after `sqlmesh audit` fails, fix the model and re-run plan/audit until clean.
Split detailed material out of SKILL.md into bundled reference files (e.g., references/model-kinds.md, references/audit-catalog.md, references/ci-templates.md) and link to them from a lean overview so the main file stays a high-level entry point.
Fix the AUDIT example syntax (`AUDIT ( name no_null_amounts, );` has an invalid trailing comma inside the directive call) so the example is copy-paste executable.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body assumes Claude's competence — it never explains what SQL, a warehouse, a cron schedule, or unit tests are — and only defines SQLMesh-specific concepts (virtual environments, breaking/non-breaking classification). Every section (Steps 1–9, Anti-patterns, Limitations) earns its place. Not a 2 because the few doc quotes are tool-specific rather than generic padding. | 3 / 3 |
Actionability | Provides copy-paste-ready, executable commands throughout ("pip install sqlmesh", "sqlmesh init duckdb", "sqlmesh plan dev", "sqlmesh run", "sqlmesh audit", "sqlmesh test") plus complete SQL model, AUDIT, YAML test, and CI examples. Not a 2 because guidance is concrete and executable rather than pseudocode or vague direction (a minor syntax slip in the AUDIT example does not undermine the overall executability). | 3 / 3 |
Workflow Clarity | The 9 steps are clearly sequenced and include a plan-review checkpoint ("The user reviews the plan, then confirms") plus CI test/plan/audit gates, but there is no explicit validate→fix→retry feedback loop for the batch/destructive operations this skill drives. It is not a 3 because the rubric specifically rewards error-recovery feedback loops for batch/database workflows, which are absent; it is not a 1 because the sequence and a verification gate are present. | 2 / 3 |
Progressive Disclosure | The body is well-organized into clear sections, but it is monolithic — all content is inline in SKILL.md with no bundled reference files (references/, scripts/, assets/ are absent) to offload detail. It is not a 3 because, for a ~230-line 9-step tool, detailed material (model-kind catalog, audit catalog, CI templates) should be split into one-level-deep reference files; it is not a 1 because sections are well-structured, not a disorganized wall of text. | 2 / 3 |
Total | 10 / 12 Passed |