Content
82%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 executable skill body: concrete commands with real flags, a numbered workflow with an explicit validation gate and failure catalog, and well-signaled references that all exist in the bundle. It loses a point per dimension to repetition across sections, an unstated fix-and-revalidate recovery loop, and contract-level detail inlined in the main file.
Suggestions
Consolidate the repeated no-target-proposals rule: state it once in 'Non-Negotiable Boundaries' and reference it from 'Run Modes' and step 6 instead of restating the full exclusion list three times.
Add an explicit error-recovery loop after step 7: instruct that when oracle_spec_guard.py fails, the agent reads <temp-dir>/specification-validation.json, fixes the specific failing condition in the candidate spec, and re-runs the compiler and guard before proceeding.
Move the durable-document 'must contain' bullet list and the readability/deduplication rules (row-length thresholds, Section 12 column schema) into references/normalized-evidence-contract.md or the specification template, keeping a short summary and link in the body so SKILL.md stays an orchestration overview.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is dense and prescriptive with no padding or explanation of concepts Claude already knows — every rule is task-specific ('Never infer a physical column from an item name alone', 'Never emit [truncated]'). It falls at anchor 4 rather than 5 because some constraints are repeated across sections: the no-target-proposals rule appears in 'Non-Negotiable Boundaries', again in 'Run Modes' ('If an older specification contains target proposals, POC assumptions, target tests...'), and again in step 6, and the durable-document bullet list overlaps the package-responsibilities list. | 4 / 5 |
Actionability | It provides fully executable commands with concrete flags for each pipeline stage — 'python <skill-dir>\scripts\oracle_spec_inventory.py <input> --module <module-id> --output <temp-manifest>', the compiler invocation with '--extraction-mode fresh --self-check', and the guard '--evidence ... --spec ... --output' — plus exact ID formats, output folder layouts, a concrete Section 12 column schema, hard row-length thresholds (4,000/5,000 characters), and a copy-paste screenshot link example. This matches 'Fully executable; copy-paste ready code or commands; specific examples cover the common cases' (fresh, refresh, and audit-only variants are all addressed). | 5 / 5 |
Workflow Clarity | The 7-step workflow has a clear sequence and an explicit validation step with commands and an exhaustive failure catalog ('Validation must fail for missing template sections or markers, ... an omitted Forms item, ... oversized Section 6/12 rows'), plus a decode-retry loop ('Repeatedly decode Forms XML text until stable'). It sits at 4 rather than 5 because there is no explicit error-recovery loop — nothing says what to do when the guard fails (fix and re-run is implied by the failure catalog, not instructed), which anchor 5 requires ('feedback loops for error recovery'). | 4 / 5 |
Progressive Disclosure | References are well signaled and genuinely one level deep: the 'Required References' section lists six files, all of which exist in references/, and the template/contract/gates/lenses detail is correctly deferred to them ('The master specification must implement all 22 numbered sections and Appendices A-J in references/specification-template.md'). It is a 4 rather than 5 because the body still inlines exhaustive policy detail (the 15-bullet durable-document list, readability/deduplication rules, row-length limits) that reads as contract material rather than the 'clear overview' anchor 5 describes. | 4 / 5 |
Total | 17 / 20 Passed |