Content
65%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 dense, domain-expert body with concrete thresholds, executable data-access code, and a valuable data-source caveat, but it is a monolithic single-file reference with no progressive disclosure and no explicitly sequenced workflow. A few time-sensitive hardcoded dates and spec-only composite scoring slightly reduce actionability and conciseness.
Suggestions
Split domain reference material (sector allocation tables, AH premium interpretation, signal frameworks) into files under references/ and keep SKILL.md as a concise overview with clearly signaled one-level-deep links.
Add an explicit step-by-step workflow section (fetch flows → compute cumulative trends → score composite → render output) with sanity checks (e.g., validating flow magnitudes against daily quotas).
Replace hardcoded dates in code examples with parameterized or relative-date placeholders, and fully implement the composite connect_score computation instead of leaving it as comment-only rules.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is mostly dense, non-obvious domain signal material (thresholds like "Quota utilization >50% = strong directional signal", "cum_20d > 30 and cum_5d > 10", and the Tushare cumulative-field data caveat), so most tokens earn their place. Minor penalties: hardcoded time-sensitive dates in examples ("start 2025-01-01", "20260101") that will go stale, and light exposition in the "Why Northbound matters" sections. This fits the 'efficient with minor trimmable instances' anchor rather than the fully lean one. | 4 / 5 |
Actionability | Mostly executable guidance: runnable signal functions, concrete Tushare/yfinance calls with documented columns, the get_southbound_flow tool, and a copy-paste output template. Minor gaps keep it below 5 — the composite connect_score dict is initialized to zeros with scoring rules only in comments, and no sample output or complete implementation of the composite is shown. | 4 / 5 |
Workflow Clarity | The sections (concepts → data access → output format) imply a fetch → compute → score → render workflow, but no explicitly sequenced procedure ties them together and no validation/sanity checkpoints exist (e.g., verifying flow data against quota bounds). This is a read-only analysis skill so the destructive-operation cap doesn't apply; sequence is implicit rather than explicit, matching the middle anchor. | 3 / 5 |
Progressive Disclosure | There are no bundle files at all — sector allocation tables, AH premium interpretation, the signal frameworks, and the output template are all inlined in a single ~255-line SKILL.md. Section organization is good, but content that clearly belongs in separate reference files is inline, and the simple-skill exception (under 50 lines) does not apply. | 3 / 5 |
Total | 14 / 20 Passed |