Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Systematically extract performance data and conditions from papers — 30 methods, 150 data points, 40 web searches budget
.claude/skills/yogsoth-ai-performance-extraction/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-06 | ✗→✓ | ▲ Improved | 103% | 0% |
| case-19 | ✗→✓ | ▲ Improved | 372% | 0% |
| case-21 | ✓→✗ | ▼ Worse | -39% | 0% |
| case-22 | ✓→✗ | ▼ Worse | 140% | 0% |
| case-20 | ✓→✓ | ＝ Same ✓ | 115% | 0% |
Extract structured performance data from papers, leaderboards, and reproducibility studies. Each data point is a (Task, Dataset, Metric, Score, Conditions) tuple with full provenance. Prioritizes primary sources (original papers) but cross-references against leaderboards and third-party reproductions.
| Resource | Floor | Target | |----------|-------|--------| | Methods covered | 20 | 30 | | Data points extracted | 100 | 150 | | Web searches | 25 | 40 | | Papers read | 15 | 30 |
<HARD-GATE>
| Metric | Current | Target | Status |
|--------|---------|--------|--------|
| Methods covered | 0 | 30 | BLOCKED |
| Data points extracted | 0 | 150 | BLOCKED |
| Web searches used | 0 | 40 | — |
| Papers read | 0 | 30 | — |
| Datasets covered | 0 | 5 | — |
| Metrics tracked | 0 | 3 | — |
</HARD-GATE>Cannot exit until data_points >= 120 (80% of target).
json{ "data_points": [ { "method": "string", "task": "string", "dataset": "string", "split": "test|val|dev", "metric": "string", "score": 0.0, "confidence_interval": [0.0, 0.0], "conditions": { "hardware": "string", "training_data_size": "string", "hyperparams_reported": true, "seeds_reported": true, "compute_budget": "string" }, "provenance": { "paper_id": "string", "table_or_figure": "string", "is_primary_source": true } } ], "coverage_summary": { "methods_covered": 0, "datasets_covered": 0, "metrics_tracked": [], "missing_data_flags": [] } }
<!-- BEGIN available-tables (generated) -->
Optional, no fixed order; the final leaf is always a sop.
| Tactic | When to use | | --- | --- | | leaderboard-harvesting | Systematically collect performance data from platforms and papers |
Optional, no fixed order; the final leaf is always a sop.
| SOP | When to use | | --- | --- | | condition-cataloging | Record evaluation conditions (data splits, hyperparams, hardware, seeds) from a paper | | score-extraction | Extract (Task, Dataset, Metric, Score, Conditions) tuples from a paper |
<!-- 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 | 59,756 | 18,605 | -69% | 1 | 1 | 0% | 8,307 | 2,275 | -73% | 0 | 0 | — |
case-02 | fail→fail | 36,815 | 28,387 | -23% | 1 | 1 | 0% | 7,451 | 3,314 | -56% | 0 | 0 | — |
case-03 | fail→fail | 41,585 | 23,460 | -44% | 1 | 1 | 0% | 8,190 | 2,488 | -70% | 0 | 0 | — |
case-04 | fail→fail | 27,048 | 10,648 | -61% | 1 | 1 | 0% | 4,326 | 1,539 | -64% | 0 | 0 | — |
case-05 | fail→fail | 19,455 | 11,566 | -41% | 1 | 1 | 0% | 2,692 | 1,975 | -27% | 0 | 0 | — |
case-06 | fail→pass | 19,307 | 55,106 | +185% | 1 | 1 | 0% | 4,454 | 9,057 | +103% | 0 | 0 | — |
case-07 | fail→fail | 20,305 | 12,907 | -36% | 1 | 1 | 0% | 2,554 | 2,232 | -13% | 0 | 0 | — |
case-08 | fail→fail | 10,939 | 6,971 | -36% | 1 | 1 | 0% | 2,291 | 1,869 | -18% | 0 | 0 | — |
case-09 | fail→fail | 21,146 | 45,509 | +115% | 1 | 1 | 0% | 4,631 | 9,059 | +96% | 0 | 0 | — |
case-10 | fail→fail | 36,941 | 7,798 | -79% | 1 | 1 | 0% | 7,566 | 1,520 | -80% | 0 | 0 | — |
case-11 | fail→fail | 7,557 | 42,984 | +469% | 1 | 1 | 0% | 1,501 | 9,059 | +504% | 0 | 0 | — |
case-12 | fail→fail | 28,477 | 13,946 | -51% | 1 | 1 | 0% | 6,726 | 2,013 | -70% | 0 | 0 | — |
case-13 | fail→fail | 36,652 | 38,835 | +6% | 1 | 1 | 0% | 4,235 | 9,054 | +114% | 0 | 0 | — |
case-14 | fail→fail | 31,750 | 9,528 | -70% | 1 | 1 | 0% | 3,870 | 1,394 | -64% | 0 | 0 | — |
case-15 | fail→fail | 17,995 | 15,350 | -15% | 1 | 1 | 0% | 3,605 | 1,770 | -51% | 0 | 0 | — |
case-16 | fail→fail | 17,735 | 8,693 | -51% | 1 | 1 | 0% | 3,314 | 1,575 | -52% | 0 | 0 | — |
case-17 | fail→fail | 14,639 | 11,679 | -20% | 1 | 1 | 0% | 2,673 | 1,523 | -43% | 0 | 0 | — |
case-18 | fail→fail | 7,905 | 12,350 | +56% | 1 | 1 | 0% | 1,615 | 1,340 | -17% | 0 | 0 | — |
case-19 | fail→pass | 11,689 | 36,569 | +213% | 1 | 1 | 0% | 1,919 | 9,053 | +372% | 0 | 0 | — |
case-20 | pass→pass | 14,342 | 26,218 | +83% | 1 | 1 | 0% | 2,145 | 4,621 | +115% | 0 | 0 | — |
case-21 | pass→fail | 37,141 | 27,636 | -26% | 1 | 1 | 0% | 4,348 | 2,643 | -39% | 0 | 0 | — |
case-22 | pass→fail | 16,395 | 34,228 | +109% | 1 | 1 | 0% | 3,631 | 8,729 | +140% | 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 9 counted toward the lift figure. The other 13 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 0 percentage points is the difference between those two pass rates over the 9 comparable cases. 4 cases got worse with the skill loaded, and they are 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.