Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Use when writing Java code with `dev.axllm:ax` for GEPA, Pareto tradeoffs, reflection clients, metric budgets, optimizer state, and artifacts.
.claude/skills/ax-llm-ax-java-gepa/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-01 | ✗→✓ | ▲ Improved | 16% | 0% |
| case-03 | ✗→✓ | ▲ Improved | -1% | 0% |
| case-07 | ✗→✓ | ▲ Improved | -62% | 0% |
| case-10 | ✗→✓ | ▲ Improved | -73% | 0% |
| case-11 | ✗→✓ | ▲ Improved | -37% | 0% |
This skill helps an agent write Java code with the generated Ax package dev.axllm:ax. Use the generated package API, examples, and manifests; do not import TypeScript-only APIs unless you are editing the TypeScript package.
dev.axllm:ax.API.md and axir-api.json.axir-capabilities.json.examples/.javascript-quickjs, python-pyodide.javaAxGEPA engine = new AxGEPA(reflectionClient, java.util.Map.of()); var result = engine.optimize(request, evaluator);
Ax.optimize, Ax.playbook, AxPlaybook, AxBootstrapFewShot, AxGEPA, OptimizerEngine, OptimizerEvaluatorprovider-api examples only when the user explicitly has provider credentials available.no-key examples for deterministic local checks and provider request mapping.tools/*/skills/ into user packages.| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→pass | 11,160 | 9,674 | -13% | 1 | 1 | 0% | 2,200 | 2,544 | +16% | 0 | 0 | — |
case-02 | fail→fail | 34,453 | 9,629 | -72% | 1 | 1 | 0% | 1,835 | 2,319 | +26% | 0 | 0 | — |
case-03 | fail→pass | 9,618 | 7,279 | -24% | 1 | 1 | 0% | 1,993 | 1,978 | -1% | 0 | 0 | — |
case-04 | pass→pass | 10,315 | 7,204 | -30% | 1 | 1 | 0% | 2,085 | 1,797 | -14% | 0 | 0 | — |
case-05 | pass→pass | 9,229 | 7,275 | -21% | 1 | 1 | 0% | 1,966 | 2,004 | +2% | 0 | 0 | — |
case-06 | pass→pass | 31,695 | 7,878 | -75% | 1 | 1 | 0% | 3,458 | 1,939 | -44% | 0 | 0 | — |
case-07 | fail→pass | 11,884 | 2,703 | -77% | 1 | 1 | 0% | 2,262 | 850 | -62% | 0 | 0 | — |
case-08 | pass→pass | 12,720 | 3,579 | -72% | 1 | 1 | 0% | 2,428 | 1,012 | -58% | 0 | 0 | — |
case-09 | pass→pass | 8,596 | 2,305 | -73% | 1 | 1 | 0% | 1,599 | 810 | -49% | 0 | 0 | — |
case-10 | fail→pass | 11,309 | 1,278 | -89% | 1 | 1 | 0% | 2,006 | 539 | -73% | 0 | 0 | — |
case-11 | fail→pass | 8,762 | 3,058 | -65% | 1 | 1 | 0% | 1,603 | 1,009 | -37% | 0 | 0 | — |
case-12 | fail→pass | 9,971 | 2,829 | -72% | 1 | 1 | 0% | 1,779 | 918 | -48% | 0 | 0 | — |
case-13 | fail→pass | 13,895 | 1,864 | -87% | 1 | 1 | 0% | 2,411 | 713 | -70% | 0 | 0 | — |
case-14 | fail→pass | 10,022 | 3,887 | -61% | 1 | 1 | 0% | 1,987 | 1,156 | -42% | 0 | 0 | — |
case-15 | pass→pass | 13,571 | 5,995 | -56% | 1 | 1 | 0% | 2,334 | 1,453 | -38% | 0 | 0 | — |
case-16 | pass→pass | 6,242 | 1,263 | -80% | 1 | 1 | 0% | 1,100 | 597 | -46% | 0 | 0 | — |
case-17 | fail→pass | 10,590 | 1,654 | -84% | 1 | 1 | 0% | 2,236 | 593 | -73% | 0 | 0 | — |
case-18 | fail→pass | 14,002 | 1,598 | -89% | 1 | 1 | 0% | 2,508 | 613 | -76% | 0 | 0 | — |
case-19 | pass→pass | 12,882 | 7,224 | -44% | 1 | 1 | 0% | 2,162 | 1,640 | -24% | 0 | 0 | — |
case-20 | fail→pass | 18,295 | 948 | -95% | 1 | 1 | 0% | 3,299 | 547 | -83% | 0 | 0 | — |
case-21 | pass→pass | 8,364 | 1,622 | -81% | 1 | 1 | 0% | 1,471 | 619 | -58% | 0 | 0 | — |
case-22 | pass→pass | 7,900 | 1,871 | -76% | 1 | 1 | 0% | 1,536 | 702 | -54% | 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 21 counted toward the lift figure. The other 1 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 +50 percentage points is the difference between those two pass rates over the 21 comparable cases.
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.