Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Apply Miller's Law — chunk information into groups of ~4 to work within working memory limits.
.claude/skills/owl-listener-millers-law/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-02 | ✗→✓ | ▲ Improved | 42% | 0% |
| case-14 | ✗→✓ | ▲ Improved | 14% | 0% |
| case-17 | ✗→✓ | ▲ Improved | 9% | 0% |
| case-10 | ✓→✗ | ▼ Worse | 19% | 0% |
| case-11 | ✓→✗ | ▼ Worse | 33% | 0% |
You are an expert in cognitive psychology as it applies to information design and interface structure.
You apply chunking and grouping strategies informed by working memory research to make interfaces easier to scan, understand, and recall.
George Miller's 1956 paper proposed that working memory can hold 7 ± 2 items (5–9). This figure has been widely cited in UX design — and just as widely misapplied. More recent research (particularly Nelson Cowan, 2001) suggests the realistic limit for meaningful chunks in working memory is closer to 4 ± 1. The important nuance Miller himself made: the "7" applies to chunks, not raw items. A chunk is whatever unit has meaning to the person — a word, a concept, a familiar pattern. What this means for design:
555-867-5309, XXXX-XXXX verification codes) for easier recall| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | pass→pass | 14,147 | 11,642 | -18% | 1 | 1 | 0% | 2,334 | 2,558 | +10% | 0 | 0 | — |
case-02 | fail→pass | 11,338 | 13,250 | +17% | 1 | 1 | 0% | 1,960 | 2,789 | +42% | 0 | 0 | — |
case-03 | pass→pass | 13,412 | 12,893 | -4% | 1 | 1 | 0% | 2,239 | 2,552 | +14% | 0 | 0 | — |
case-04 | pass→pass | 12,540 | 11,303 | -10% | 1 | 1 | 0% | 2,284 | 2,532 | +11% | 0 | 0 | — |
case-09 | pass→pass | 14,891 | 15,513 | +4% | 1 | 1 | 0% | 2,509 | 3,063 | +22% | 0 | 0 | — |
case-05 | pass→pass | 11,959 | 10,767 | -10% | 1 | 1 | 0% | 2,086 | 2,493 | +20% | 0 | 0 | — |
case-06 | pass→pass | 10,027 | 5,973 | -40% | 1 | 1 | 0% | 1,640 | 1,571 | -4% | 0 | 0 | — |
case-07 | pass→pass | 14,626 | 15,767 | +8% | 1 | 1 | 0% | 2,256 | 2,975 | +32% | 0 | 0 | — |
case-08 | pass→pass | 13,985 | 11,734 | -16% | 1 | 1 | 0% | 2,237 | 2,461 | +10% | 0 | 0 | — |
case-10 | pass→fail | 13,472 | 14,003 | +4% | 1 | 1 | 0% | 2,372 | 2,825 | +19% | 0 | 0 | — |
case-11 | pass→fail | 11,225 | 12,060 | +7% | 1 | 1 | 0% | 1,852 | 2,454 | +33% | 0 | 0 | — |
case-12 | pass→pass | 9,428 | 9,328 | -1% | 1 | 1 | 0% | 1,828 | 2,264 | +24% | 0 | 0 | — |
case-13 | pass→pass | 11,002 | 11,946 | +9% | 1 | 1 | 0% | 1,889 | 2,420 | +28% | 0 | 0 | — |
case-14 | fail→pass | 14,471 | 14,567 | +1% | 1 | 1 | 0% | 2,456 | 2,800 | +14% | 0 | 0 | — |
case-15 | pass→pass | 11,836 | 13,907 | +17% | 1 | 1 | 0% | 1,989 | 2,839 | +43% | 0 | 0 | — |
case-16 | pass→pass | 8,647 | 10,964 | +27% | 1 | 1 | 0% | 1,466 | 2,292 | +56% | 0 | 0 | — |
case-17 | fail→pass | 13,016 | 11,770 | -10% | 1 | 1 | 0% | 2,371 | 2,579 | +9% | 0 | 0 | — |
case-18 | pass→pass | 10,697 | 12,973 | +21% | 1 | 1 | 0% | 2,012 | 2,753 | +37% | 0 | 0 | — |
case-19 | pass→pass | 14,208 | 11,918 | -16% | 1 | 1 | 0% | 2,362 | 2,563 | +9% | 0 | 0 | — |
case-20 | pass→pass | 10,963 | 10,915 | -0% | 1 | 1 | 0% | 2,048 | 2,506 | +22% | 0 | 0 | — |
case-21 | pass→pass | 12,382 | 13,584 | +10% | 1 | 1 | 0% | 1,909 | 2,650 | +39% | 0 | 0 | — |
case-22 | pass→pass | 12,543 | 12,302 | -2% | 1 | 1 | 0% | 2,043 | 2,524 | +24% | 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. 2 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.