Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Self-improving prompts through meta-level optimization
.claude/skills/miosa-osa-meta-prompting/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-01 | ✗→✓ | ▲ Improved | 70% | 0% |
| case-03 | ✗→✓ | ▲ Improved | 25% | 0% |
| case-04 | ✗→✓ | ▲ Improved | 37% | 0% |
| case-08 | ✗→✓ | ▲ Improved | 17% | 0% |
| case-14 | ✗→✓ | ▲ Improved | -64% | 0% |
> Optimize prompts through iterative self-improvement.
/meta-prompting "<prompt>" [--iterations <n>] [--target <score>]Takes a prompt, executes it, scores the result, then uses the LLM to optimize the prompt for better results. Iterates until the quality target is met or max iterations reached. Stores optimized prompts in the skill library for reuse.
Optimization dimensions: clarity enhancement, example addition, instruction refinement, context optimization.
bash# Optimize a code generation prompt /meta-prompting "Generate a rate limiter in Go" --target 8 # Optimize a content prompt with more iterations /meta-prompting "Write a pitch for enterprise clients" --iterations 5 # Optimize a system prompt /meta-prompting "You are a security auditor..." --target 9
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→pass | 23,141 | 36,196 | +56% | 1 | 1 | 0% | 4,435 | 7,549 | +70% | 0 | 0 | — |
case-02 | pass→pass | 16,749 | 30,110 | +80% | 1 | 1 | 0% | 2,587 | 5,745 | +122% | 0 | 0 | — |
case-03 | fail→pass | 23,768 | 28,763 | +21% | 1 | 1 | 0% | 3,815 | 4,767 | +25% | 0 | 0 | — |
case-04 | fail→pass | 16,425 | 18,764 | +14% | 1 | 1 | 0% | 2,977 | 4,073 | +37% | 0 | 0 | — |
case-05 | pass→pass | 15,978 | 19,533 | +22% | 1 | 1 | 0% | 2,342 | 4,088 | +75% | 0 | 0 | — |
case-06 | pass→pass | 15,075 | 22,130 | +47% | 1 | 1 | 0% | 2,679 | 4,865 | +82% | 0 | 0 | — |
case-07 | pass→pass | 13,004 | 20,956 | +61% | 1 | 1 | 0% | 2,039 | 4,495 | +120% | 0 | 0 | — |
case-08 | fail→pass | 11,127 | 10,531 | -5% | 1 | 1 | 0% | 1,863 | 2,175 | +17% | 0 | 0 | — |
case-09 | pass→pass | 8,769 | 3,482 | -60% | 1 | 1 | 0% | 1,452 | 899 | -38% | 0 | 0 | — |
case-10 | pass→pass | 9,235 | 3,309 | -64% | 1 | 1 | 0% | 1,542 | 835 | -46% | 0 | 0 | — |
case-11 | pass→pass | 8,823 | 5,722 | -35% | 1 | 1 | 0% | 1,459 | 1,339 | -8% | 0 | 0 | — |
case-12 | pass→pass | 8,499 | 9,293 | +9% | 1 | 1 | 0% | 1,518 | 1,908 | +26% | 0 | 0 | — |
case-13 | pass→pass | 5,820 | 4,127 | -29% | 1 | 1 | 0% | 948 | 982 | +4% | 0 | 0 | — |
case-14 | fail→pass | 10,264 | 1,705 | -83% | 1 | 1 | 0% | 1,588 | 573 | -64% | 0 | 0 | — |
case-15 | fail→pass | 7,846 | 2,511 | -68% | 1 | 1 | 0% | 1,239 | 773 | -38% | 0 | 0 | — |
case-16 | pass→pass | 8,763 | 2,924 | -67% | 1 | 1 | 0% | 1,490 | 692 | -54% | 0 | 0 | — |
case-17 | pass→pass | 5,160 | 3,126 | -39% | 1 | 1 | 0% | 839 | 878 | +5% | 0 | 0 | — |
case-18 | pass→pass | 7,285 | 5,467 | -25% | 1 | 1 | 0% | 1,138 | 1,202 | +6% | 0 | 0 | — |
case-19 | pass→pass | 2,720 | 3,262 | +20% | 1 | 1 | 0% | 429 | 810 | +89% | 0 | 0 | — |
case-20 | pass→pass | 5,330 | 4,804 | -10% | 1 | 1 | 0% | 950 | 1,145 | +21% | 0 | 0 | — |
case-21 | pass→pass | 6,436 | 6,890 | +7% | 1 | 1 | 0% | 1,328 | 1,697 | +28% | 0 | 0 | — |
case-22 | pass→pass | 4,982 | 2,924 | -41% | 1 | 1 | 0% | 848 | 808 | -5% | 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 +27 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.