Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Strategy: Treat the gap set as an investment portfolio — use risk/return/diversity optimization to select the optimal gap portfolio
.claude/skills/yogsoth-ai-hypothesis-formation-portfolio-optimization/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-02 | ✗→✓ | ▲ Improved | 14% | 0% |
| case-03 | ✗→✓ | ▲ Improved | 14% | 0% |
| case-08 | ✗→✓ | ▲ Improved | 31% | 0% |
| case-10 | ✗→✓ | ▲ Improved | -14% | 0% |
| case-14 | ✗→✓ | ▲ Improved | -10% | 0% |
Treat the set of research gaps as an investment portfolio: not only assess the value of each individual gap, but also analyze the correlation and complementarity among gaps, selecting the portfolio with the greatest overall return and the most diversified risk.
Core principle: The value of an individual gap is not equal to its marginal contribution within the portfolio.
Draw an analogy between gap portfolio optimization and Markowitz portfolio theory:
Return: the expected academic/applied value of the gap (importance × impact)
Risk: the uncertainty of the gap (reciprocal of feasibility + reciprocal of technical maturity)
Correlation: Do two gaps depend on the same methods, data, or prerequisite work? Highly correlated gaps have a higher probability of failing simultaneously and should receive lower portfolio weight.
Diversity: The portfolio should span different subfields, different methodologies, and different time horizons (short-term deliverable + long-term breakthrough).
Efficient Frontier: At a given risk level, find the gap portfolio with the greatest return; or at a given return target, find the portfolio with the least risk.
Key insight: A medium-value gap with low correlation to other gaps may be more worth including in the portfolio than a high-value but highly correlated gap — because it provides genuine diversification.
| Tier | Number of gaps | Correlation analysis | Portfolio size | Final output | |------|---------|-----------|---------|---------| | S | 20–30 | Qualitative (high/medium/low) | Select 3–5 | Recommended portfolio + rationale | | M | 31–50 | Semi-quantitative (correlation matrix) | Select 5–8 | Recommended portfolio + efficient frontier chart + alternative portfolios | | L | 50+ | Quantitative (method/data overlap) | Select 8–12 | Full portfolio analysis + efficient frontier + risk decomposition |
gap-normalization SOP: unify gap format, extract methodology labels and data dependenciesimportance-scoring, feasibility-scoring, novelty-scoring, impact-scoring)scoring-matrix-construction tactic: build the scoring matrixpriority-sensitivity-testing tactic: test portfolio robustness under different risk preferencespriority-synthesis SOP: output the recommended portfolio + efficient frontier analysisRecord after each round:
<!-- BEGIN available-tables (generated) -->
Optional, no fixed order; the final leaf is always a sop.
| Tactic | When to use | | --- | --- | | hypothesis-formation-scoring-matrix-construction | Tactic: orchestrate multi-dimensional scoring SOPs to build a comprehensive assessment matrix for all gaps | | priority-sensitivity-testing | Tactic: perturb scoring weights to test the robustness of the gap ranking against weight choice |
Optional, no fixed order; the final leaf is always a sop.
| SOP | When to use | | --- | --- | | gap-normalization | SOP: Unify gaps from different sources into the standard GapRecord format |
<!-- END available-tables (generated) -->
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→fail | 33,613 | 50,241 | +49% | 1 | 1 | 0% | 5,100 | 7,115 | +40% | 0 | 0 | — |
case-02 | fail→pass | 73,791 | 35,112 | -52% | 1 | 1 | 0% | 6,229 | 7,085 | +14% | 0 | 0 | — |
case-03 | fail→pass | 39,187 | 40,199 | +3% | 1 | 1 | 0% | 6,224 | 7,115 | +14% | 0 | 0 | — |
case-04 | pass→pass | 9,204 | 10,329 | +12% | 1 | 1 | 0% | 1,402 | 2,407 | +72% | 0 | 0 | — |
case-05 | pass→pass | 14,130 | 11,588 | -18% | 1 | 1 | 0% | 2,663 | 2,822 | +6% | 0 | 0 | — |
case-06 | pass→pass | 9,787 | 10,398 | +6% | 1 | 1 | 0% | 1,518 | 2,374 | +56% | 0 | 0 | — |
case-11 | pass→pass | 15,435 | 14,825 | -4% | 1 | 1 | 0% | 2,280 | 3,151 | +38% | 0 | 0 | — |
case-07 | pass→fail | 18,178 | 14,457 | -20% | 1 | 1 | 0% | 2,835 | 3,431 | +21% | 0 | 0 | — |
case-08 | fail→pass | 13,321 | 13,052 | -2% | 1 | 1 | 0% | 2,375 | 3,111 | +31% | 0 | 0 | — |
case-09 | pass→pass | 9,432 | 9,207 | -2% | 1 | 1 | 0% | 1,480 | 2,326 | +57% | 0 | 0 | — |
case-10 | fail→pass | 16,190 | 6,913 | -57% | 1 | 1 | 0% | 2,388 | 2,051 | -14% | 0 | 0 | — |
case-12 | pass→pass | 20,827 | 15,565 | -25% | 1 | 1 | 0% | 2,987 | 3,250 | +9% | 0 | 0 | — |
case-13 | pass→pass | 13,045 | 9,368 | -28% | 1 | 1 | 0% | 1,901 | 2,319 | +22% | 0 | 0 | — |
case-14 | fail→pass | 11,367 | 4,009 | -65% | 1 | 1 | 0% | 1,651 | 1,485 | -10% | 0 | 0 | — |
case-15 | pass→pass | 18,862 | 6,026 | -68% | 1 | 1 | 0% | 2,987 | 1,876 | -37% | 0 | 0 | — |
case-16 | fail→pass | 23,721 | 3,002 | -87% | 1 | 1 | 0% | 1,856 | 1,343 | -28% | 0 | 0 | — |
case-17 | fail→pass | 25,005 | 3,178 | -87% | 1 | 1 | 0% | 1,323 | 1,337 | +1% | 0 | 0 | — |
case-18 | fail→pass | 25,093 | 3,519 | -86% | 1 | 1 | 0% | 1,827 | 1,412 | -23% | 0 | 0 | — |
case-19 | pass→pass | 15,167 | 14,711 | -3% | 1 | 1 | 0% | 2,272 | 3,167 | +39% | 0 | 0 | — |
case-20 | pass→pass | 15,483 | 13,373 | -14% | 1 | 1 | 0% | 2,204 | 2,646 | +20% | 0 | 0 | — |
case-21 | pass→pass | 11,137 | 11,599 | +4% | 1 | 1 | 0% | 1,651 | 2,622 | +59% | 0 | 0 | — |
case-22 | pass→pass | 16,437 | 6,390 | -61% | 1 | 1 | 0% | 2,387 | 1,807 | -24% | 0 | 0 | — |
DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 22 cases were attempted, and 20 counted toward the lift figure. The other 2 produced results that are not comparable between the two arms, so they are excluded from the headline rather than averaged into it. The headline lift of +32 percentage points is the difference between those two pass rates over the 20 comparable cases. 1 case got worse with the skill loaded, and it is included in that figure.
Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.
Other measured skills in the registry, with their headline benchmark lift.