Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Systematic improvement of existing agents through performance analysis, prompt engineering, and continuous iteration.
.claude/skills/dokhacgiakhoa-agent-orchestration-improve-agent/SKILL.md| Model | Eval pass | Runs |
|---|---|---|
| gemini-3.6-flash | 100% | 31 |
| gemini-3.1-pro-preview | 100% | 3 |
| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-18 | ✗→✓ | ▲ Improved | 25% | 0% |
| case-06 | ✓→✓ | ＝ Same ✓ | 20% | 0% |
| case-01 | ✓→✓ | ＝ Same ✓ | 22% | 0% |
| case-02 | ✓→✓ | ＝ Same ✓ | 21% | 0% |
| case-03 | ✓→✓ | ＝ Same ✓ | 68% | 0% |
Systematic improvement of existing agents through performance analysis, prompt engineering, and continuous iteration.
Extended thinking: Agent optimization requires a data-driven approach combining performance metrics, user feedback analysis, and advanced prompt engineering techniques. Success depends on systematic evaluation, targeted improvements, and rigorous testing with rollback capabilities for production safety.]
Comprehensive analysis of agent performance using context-manager for historical data collection.
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-06 | pass→pass | 13,088 | 13,150 | +0% | 1 | 1 | 0% | 2,358 | 2,841 | +20% | 0 | 0 | — |
case-01 | pass→pass | 13,271 | 11,383 | -14% | 1 | 1 | 0% | 1,977 | 2,403 | +22% | 0 | 0 | — |
case-02 | pass→pass | 16,921 | 15,532 | -8% | 1 | 1 | 0% | 2,553 | 3,080 | +21% | 0 | 0 | — |
case-03 | pass→pass | 8,151 | 9,783 | +20% | 1 | 1 | 0% | 1,394 | 2,346 | +68% | 0 | 0 | — |
case-04 | pass→pass | 6,753 | 8,315 | +23% | 1 | 1 | 0% | 999 | 2,011 | +101% | 0 | 0 | — |
case-05 | pass→pass | 13,458 | 14,102 | +5% | 1 | 1 | 0% | 2,128 | 3,031 | +42% | 0 | 0 | — |
case-07 | pass→pass | 9,852 | 11,530 | +17% | 1 | 1 | 0% | 1,719 | 2,867 | +67% | 0 | 0 | — |
case-08 | pass→pass | 17,772 | 15,419 | -13% | 1 | 1 | 0% | 3,362 | 3,717 | +11% | 0 | 0 | — |
case-09 | pass→pass | 14,043 | 15,603 | +11% | 1 | 1 | 0% | 2,517 | 3,066 | +22% | 0 | 0 | — |
case-10 | pass→pass | 11,416 | 11,498 | +1% | 1 | 1 | 0% | 1,673 | 2,967 | +77% | 0 | 0 | — |
case-11 | pass→pass | 11,949 | 9,191 | -23% | 1 | 1 | 0% | 1,812 | 2,479 | +37% | 0 | 0 | — |
case-12 | pass→pass | 9,551 | 11,142 | +17% | 1 | 1 | 0% | 1,629 | 2,447 | +50% | 0 | 0 | — |
case-13 | pass→pass | 14,343 | 15,267 | +6% | 1 | 1 | 0% | 2,355 | 2,939 | +25% | 0 | 0 | — |
case-14 | pass→pass | 16,146 | 19,929 | +23% | 1 | 1 | 0% | 2,707 | 4,055 | +50% | 0 | 0 | — |
case-15 | pass→pass | 10,971 | 11,276 | +3% | 1 | 1 | 0% | 1,906 | 2,748 | +44% | 0 | 0 | — |
case-16 | fail→fail | 16,465 | 18,869 | +15% | 1 | 1 | 0% | 2,941 | 4,439 | +51% | 0 | 0 | — |
case-17 | fail→fail | 10,179 | 6,720 | -34% | 1 | 1 | 0% | 1,665 | 1,927 | +16% | 0 | 0 | — |
case-18 | fail→pass | 19,226 | 19,239 | +0% | 1 | 1 | 0% | 3,413 | 4,264 | +25% | 0 | 0 | — |
case-19 | pass→pass | 20,172 | 17,046 | -15% | 1 | 1 | 0% | 3,209 | 3,699 | +15% | 0 | 0 | — |
case-20 | pass→pass | 13,826 | 8,908 | -36% | 1 | 1 | 0% | 2,630 | 2,559 | -3% | 0 | 0 | — |
case-21 | fail→fail | 11,864 | 12,004 | +1% | 1 | 1 | 0% | 2,107 | 2,926 | +39% | 0 | 0 | — |
case-22 | pass→pass | 13,982 | 14,597 | +4% | 1 | 1 | 0% | 2,434 | 3,323 | +37% | 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. The headline lift of +5 percentage points is the difference between those two pass rates over the 22 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.