Content
63%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, largely executable reference whose strongest material is the module-specific calibration knowledge (sign conventions, tail classification thresholds, reproducibility rules) that Claude could not reconstruct. It is held back by inline textbook background, a formula-only stress-testing section, and a monolithic structure with no progressive disclosure into reference files.
Suggestions
Trim or drop the textbook passages Claude already knows (VaR/CVaR definitions, method advantages/disadvantages table, skewness/kurtosis explanations, Basel note) and keep only the module-specific behavior.
Make the stress-testing workflow executable like the rest: expose or reference a function that applies STRESS_SCENARIOS shocks to positions instead of leaving a Σ formula for inline retyping.
Split the scenario tables, the EVT/GPD deep-dive, and the report template into one-level-deep reference files (e.g. references/scenarios.md, references/evt.md) linked from a leaner SKILL.md.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | Mostly efficient — sign conventions, the parametric-vs-historical calibration table, the 2σ tail_type rule, and seed discipline are genuinely non-obvious — but it also explains concepts Claude already knows: VaR/CVaR definitions, the three-method advantages/disadvantages table, skewness/kurtosis meanings, and the Basel II/III row. Not 2: the padding is embedded and brief rather than dominating; not 4: several textbook passages could be cut outright. | 3 / 5 |
Actionability | An executable import block, concrete call signatures with defaults, and documented return keys make most sections copy-paste ready (e.g. `historical_var(returns, confidence=0.95)`, `fit_gpd_tail(returns, threshold_pct=5.0)`). Not 5: the stress-testing section — a headline capability — provides only a formula (`portfolio_loss = Σ(weight_i × shock_i × position_i)`) and a scenario dict with no callable function, unlike every other section. | 4 / 5 |
Workflow Clarity | The Analysis Steps give a clear 7-step sequence with explicit checkpoints such as "If you ever compute a CVaR below its VaR, the tail mask is wrong" and "Check that shape_xi is stable across a few nearby threshold_pct values". Not 5: there is no validate→fix→retry feedback loop and some checkpoints (e.g. stress-test adequacy assessment) are left implicit; not 3: several concrete validation checks are explicitly stated. | 4 / 5 |
Progressive Disclosure | The body is well-headered and navigable, but it is a ~310-line monolith with no bundle files at all: the historical/hypothetical scenario tables, the GPD/EVT deep-dive, and the output template are inline content that would sit better in one-level-deep reference files. Not 4: a skill of this size has clear separable material; not 2: section structure is good, not minimal. | 3 / 5 |
Total | 14 / 20 Passed |