Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Analyze OSWorld-V2 agent trajectory logs and task results to produce actionable insights. Use this skill whenever the user wants to understand agent performance on OSWorld tasks — including analyzing trajectories, reviewing task results, finding error patterns, comparing code vs GUI strategies, identifying which tools/commands the agent used, or deciding which task types to scale up in the benchmark.
.claude/skills/amap-ml-analyze-traj/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-08 | ✗→✓ | ▲ Improved | -65% | 0% |
| case-13 | ✗→✓ | ▲ Improved | 19% | 0% |
| case-15 | ✗→✓ | ▲ Improved | -49% | 0% |
| case-18 | ✗→✓ | ▲ Improved | -59% | 0% |
| case-07 | ✓→✓ | ＝ Same ✓ | -44% | 0% |
If only one task is issued, analyze it directly with instruction: analyze-single-traj.md.
If multiple tasks or a whole results directory are issued, use subagents to analyze them in parallel (one agent for each task). Do not analyze them sequentially by yourself. DO NOT tell it what to do. Just ask the subagent to analyze the task in target directory and use this skill (analyze-traj) to do the analysis. Pass any user instructions to every subagent.
After the per-task reports are ready:
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→fail | 6,441 | 14,908 | +131% | 1 | 1 | 0% | 1,131 | 548 | -52% | 0 | 0 | — |
case-02 | fail→fail | 11,155 | 15,191 | +36% | 1 | 1 | 0% | 426 | 414 | -3% | 0 | 0 | — |
case-03 | fail→fail | 3,711 | 14,989 | +304% | 1 | 1 | 0% | 163 | 491 | +201% | 0 | 0 | — |
case-04 | fail→fail | 17,780 | 4,861 | -73% | 1 | 1 | 0% | 1,010 | 397 | -61% | 0 | 0 | — |
case-05 | fail→fail | 9,368 | 10,783 | +15% | 1 | 1 | 0% | 197 | 613 | +211% | 0 | 0 | — |
case-06 | fail→fail | 10,868 | 9,815 | -10% | 1 | 1 | 0% | 259 | 513 | +98% | 0 | 0 | — |
case-07 | pass→pass | 11,343 | 7,135 | -37% | 1 | 1 | 0% | 1,018 | 575 | -44% | 0 | 0 | — |
case-08 | fail→pass | 14,210 | 2,106 | -85% | 1 | 1 | 0% | 1,430 | 501 | -65% | 0 | 0 | — |
case-09 | fail→fail | 5,227 | 11,212 | +115% | 1 | 1 | 0% | 872 | 600 | -31% | 0 | 0 | — |
case-10 | fail→fail | 20,233 | 6,023 | -70% | 1 | 1 | 0% | 356 | 580 | +63% | 0 | 0 | — |
case-11 | fail→fail | 30,275 | 11,098 | -63% | 1 | 1 | 0% | 5,137 | 575 | -89% | 0 | 0 | — |
case-12 | fail→fail | 9,639 | 16,162 | +68% | 1 | 1 | 0% | 180 | 580 | +222% | 0 | 0 | — |
case-13 | fail→pass | 8,877 | 10,799 | +22% | 1 | 1 | 0% | 1,189 | 1,411 | +19% | 0 | 0 | — |
case-14 | pass→pass | 7,590 | 3,199 | -58% | 1 | 1 | 0% | 397 | 625 | +57% | 0 | 0 | — |
case-15 | fail→pass | 15,051 | 9,531 | -37% | 1 | 1 | 0% | 1,666 | 842 | -49% | 0 | 0 | — |
case-16 | pass→pass | 6,229 | 2,367 | -62% | 1 | 1 | 0% | 898 | 468 | -48% | 0 | 0 | — |
case-17 | fail→fail | 10,894 | 10,065 | -8% | 1 | 1 | 0% | 171 | 439 | +157% | 0 | 0 | — |
case-18 | fail→pass | 11,910 | 1,915 | -84% | 1 | 1 | 0% | 1,246 | 513 | -59% | 0 | 0 | — |
case-19 | fail→fail | 48,737 | 16,412 | -66% | 1 | 1 | 0% | 6,061 | 381 | -94% | 0 | 0 | — |
case-20 | pass→pass | 14,186 | 20,686 | +46% | 1 | 1 | 0% | 2,957 | 3,376 | +14% | 0 | 0 | — |
case-21 | fail→fail | 5,244 | 12,670 | +142% | 1 | 1 | 0% | 224 | 753 | +236% | 0 | 0 | — |
case-22 | fail→fail | 10,032 | 17,289 | +72% | 1 | 1 | 0% | 161 | 534 | +232% | 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 8 counted toward the lift figure. The other 14 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 +18 percentage points is the difference between those two pass rates over the 8 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.