Content
78%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 lean, well-structured body with a genuine validation loop and concrete MCP tool names — the verify step in particular is explicit and emphasized. The two real weaknesses are the broken best-practices reference and the absence of any executable code example or explicit fix-and-retry instruction.
Suggestions
Fix the broken reference: the body links to 'reference/best_practices.md' twice, but no 'reference/' directory or best_practices.md file exists in the bundle. Either ship the file (and correct the directory name if it should be 'references/') or remove the links and inline the key rules (@app.cell, no global state, reactivity).
Add an explicit error-recovery step to the Verify stage, e.g. 'If lint or runtime errors appear: read the error, fix the cell in the .py file, then re-run lint and error checks before proceeding.'
Include one minimal executable snippet (a two-cell '@app.cell' example with mo.sql or a matplotlib return) so the common notebook patterns are concrete rather than name-only.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | At ~30 lines, every line carries information Claude does not already know: tool names, sync behavior, 'No plt.show()', 'mo.sql'. There is no conceptual padding or over-explanation of what marimo is. | 5 / 5 |
Actionability | Concrete, named MCP tools ('mcp_marimo_lint_notebook', 'mcp_marimo_get_notebook_errors', etc.) and specific rules ('Use mo.sql for DuckDB queries', 'No plt.show()') give mostly executable guidance. Not 5: no actual code example (e.g. an '@app.cell' snippet or tool call with arguments), so the common cases are not copy-paste ready. | 4 / 5 |
Workflow Clarity | The Plan-Execute-Verify loop is a clear three-step sequence with an explicit, CRITICAL-flagged Verify step containing three validation checkpoints (lint, runtime errors, outputs). Not 5: there is no explicit error-recovery instruction ('if lint/errors fail, fix and re-verify') — the retry behavior is only implied by the 'Feedback Loop' title. | 4 / 5 |
Progressive Disclosure | The body is well-structured and clearly signals a one-level-deep reference via a dedicated 'Reference' section, but the referenced file 'reference/best_practices.md' does not exist in the bundle (no references/scripts/assets directories are present). The broken path means navigation fails at exactly the point detail is supposed to live, which is worse than the well-signaled structure alone would earn. | 3 / 5 |
Total | 16 / 20 Passed |