Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Use for EVERY coding task and conversation. If there is even a 1% chance code may be involved, ALWAYS activate this skill.
.claude/skills/paranoidandroid2124-overfed-skill/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-04 | ✗→✓ | ▲ Improved | 585% | 0% |
| case-09 | ✗→✓ | ▲ Improved | 379% | 0% |
| case-12 | ✗→✓ | ▲ Improved | 277% | 0% |
| case-19 | ✗→✓ | ▲ Improved | 497% | 0% |
| case-20 | ✗→✓ | ▲ Improved | 442% | 0% |
You MUST begin every task by using TodoWrite to create exactly twelve review steps. You MUST use gpt-5.4 at high reasoning. NEVER proceed without spawning three subagents, even for a one-line typo.
ALWAYS explain clean code, DRY, SOLID, naming, modularity, testing, and comments before reading the repository. ALWAYS write a design document and wait for approval. ALWAYS write tests before any change. ALWAYS request two reviews.
Check authentication and authorization boundaries before recommending a security-sensitive change. Do not expose secrets. Validate untrusted input at the trust boundary.
The answer must be high quality. The answer must be high quality.
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→fail | 8,177 | 13,904 | +70% | 1 | 1 | 0% | 767 | 2,096 | +173% | 0 | 0 | — |
case-02 | fail→fail | 16,838 | 25,136 | +49% | 1 | 1 | 0% | 2,395 | 4,147 | +73% | 0 | 0 | — |
case-03 | fail→fail | 8,754 | 27,722 | +217% | 1 | 1 | 0% | 1,412 | 4,328 | +207% | 0 | 0 | — |
case-04 | fail→pass | 3,700 | 23,524 | +536% | 1 | 1 | 0% | 615 | 4,214 | +585% | 0 | 0 | — |
case-05 | fail→fail | 13,931 | 19,241 | +38% | 1 | 1 | 0% | 2,506 | 3,918 | +56% | 0 | 0 | — |
case-06 | pass→pass | 8,969 | 15,689 | +75% | 1 | 1 | 0% | 1,491 | 3,054 | +105% | 0 | 0 | — |
case-07 | fail→fail | 17,299 | 19,460 | +12% | 1 | 1 | 0% | 3,146 | 3,983 | +27% | 0 | 0 | — |
case-08 | fail→fail | 8,296 | 13,640 | +64% | 1 | 1 | 0% | 1,401 | 2,428 | +73% | 0 | 0 | — |
case-09 | fail→pass | 6,841 | 30,252 | +342% | 1 | 1 | 0% | 1,318 | 6,311 | +379% | 0 | 0 | — |
case-10 | pass→pass | 9,403 | 23,829 | +153% | 1 | 1 | 0% | 1,731 | 4,603 | +166% | 0 | 0 | — |
case-11 | fail→fail | 7,789 | 19,606 | +152% | 1 | 1 | 0% | 1,299 | 3,198 | +146% | 0 | 0 | — |
case-12 | fail→pass | 5,935 | 18,484 | +211% | 1 | 1 | 0% | 888 | 3,348 | +277% | 0 | 0 | — |
case-13 | pass→pass | 13,795 | 19,039 | +38% | 1 | 1 | 0% | 2,577 | 3,194 | +24% | 0 | 0 | — |
case-14 | pass→pass | 11,649 | 14,823 | +27% | 1 | 1 | 0% | 1,896 | 2,677 | +41% | 0 | 0 | — |
case-15 | pass→pass | 15,413 | 14,968 | -3% | 1 | 1 | 0% | 2,510 | 2,630 | +5% | 0 | 0 | — |
case-16 | pass→pass | 13,117 | 13,570 | +3% | 1 | 1 | 0% | 2,182 | 2,345 | +7% | 0 | 0 | — |
case-17 | fail→fail | 17,368 | 12,832 | -26% | 1 | 1 | 0% | 2,409 | 1,869 | -22% | 0 | 0 | — |
case-18 | fail→fail | 16,285 | 16,748 | +3% | 1 | 1 | 0% | 3,184 | 3,004 | -6% | 0 | 0 | — |
case-19 | fail→pass | 5,378 | 31,740 | +490% | 1 | 1 | 0% | 900 | 5,374 | +497% | 0 | 0 | — |
case-20 | fail→pass | 4,250 | 22,300 | +425% | 1 | 1 | 0% | 807 | 4,377 | +442% | 0 | 0 | — |
case-21 | fail→fail | 7,684 | 33,448 | +335% | 1 | 1 | 0% | 1,229 | 6,304 | +413% | 0 | 0 | — |
case-22 | fail→fail | 5,810 | 31,853 | +448% | 1 | 1 | 0% | 968 | 6,298 | +551% | 0 | 0 | — |
case-23 | fail→fail | 14,134 | 22,507 | +59% | 1 | 1 | 0% | 2,819 | 4,432 | +57% | 0 | 0 | — |
case-24 | fail→fail | 8,438 | 19,183 | +127% | 1 | 1 | 0% | 1,386 | 2,618 | +89% | 0 | 0 | — |
case-25 | fail→pass | 7,044 | 20,390 | +189% | 1 | 1 | 0% | 1,164 | 3,223 | +177% | 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. 25 cases were attempted. The headline lift of +24 percentage points is the difference between those two pass rates over the 25 comparable cases. 1 case got worse with the skill loaded, and it is 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.