Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Save what we just did as a reusable skill. Use at the end of a session to capture a workflow, technique, or process that you want to repeat.
.claude/skills/iyadhkhalfallah-save-skill/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-03 | ✗→✓ | ▲ Improved | -37% | 0% |
| case-01 | ✗→✓ | ▲ Improved | -10% | 0% |
| case-02 | ✗→✓ | ▲ Improved | -27% | 0% |
| case-07 | ✗→✓ | ▲ Improved | -18% | 0% |
| case-11 | ✗→✓ | ▲ Improved | -13% | 0% |
The user wants to save what was done in this session as a reusable Claude Code skill.
disable-model-invocation: true if user-only$ARGUMENTS for any variable parts (file names, branch names, etc.)/skill-name.If the session involved setting up a new API endpoint:
.claude/skills/new-endpoint/SKILL.mdyaml--- name: new-endpoint description: Scaffold a new REST API endpoint with tests and documentation disable-model-invocation: true --- Create a new API endpoint for $ARGUMENTS: 1. Create the route handler in src/routes/ 2. Add request/response validation schemas 3. Write integration tests 4. Update the API documentation 5. Run tests to verify
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-03 | fail→pass | 12,051 | 4,395 | -64% | 1 | 1 | 0% | 1,903 | 1,204 | -37% | 0 | 0 | — |
case-01 | fail→pass | 10,375 | 6,213 | -40% | 1 | 1 | 0% | 1,755 | 1,581 | -10% | 0 | 0 | — |
case-02 | fail→pass | 10,832 | 5,305 | -51% | 1 | 1 | 0% | 1,870 | 1,373 | -27% | 0 | 0 | — |
case-04 | pass→fail | 5,855 | 4,204 | -28% | 1 | 1 | 0% | 1,103 | 1,219 | +11% | 0 | 0 | — |
case-05 | pass→pass | 8,993 | 5,045 | -44% | 1 | 1 | 0% | 1,478 | 1,220 | -17% | 0 | 0 | — |
case-06 | pass→fail | 5,632 | 5,309 | -6% | 1 | 1 | 0% | 935 | 787 | -16% | 0 | 0 | — |
case-07 | fail→pass | 7,737 | 6,261 | -19% | 1 | 1 | 0% | 1,250 | 1,021 | -18% | 0 | 0 | — |
case-08 | pass→pass | 8,590 | 2,441 | -72% | 1 | 1 | 0% | 1,437 | 919 | -36% | 0 | 0 | — |
case-09 | fail→fail | 12,054 | 6,094 | -49% | 1 | 1 | 0% | 1,861 | 828 | -56% | 0 | 0 | — |
case-10 | pass→pass | 6,582 | 2,581 | -61% | 1 | 1 | 0% | 990 | 888 | -10% | 0 | 0 | — |
case-11 | fail→pass | 10,696 | 6,717 | -37% | 1 | 1 | 0% | 1,817 | 1,584 | -13% | 0 | 0 | — |
case-12 | pass→pass | 7,818 | 5,955 | -24% | 1 | 1 | 0% | 1,465 | 1,460 | -0% | 0 | 0 | — |
case-13 | pass→pass | 7,016 | 3,194 | -54% | 1 | 1 | 0% | 1,118 | 967 | -14% | 0 | 0 | — |
case-14 | pass→pass | 14,595 | 10,191 | -30% | 1 | 1 | 0% | 2,286 | 2,232 | -2% | 0 | 0 | — |
case-15 | fail→pass | 8,318 | 3,389 | -59% | 1 | 1 | 0% | 1,356 | 1,015 | -25% | 0 | 0 | — |
case-16 | pass→pass | 11,334 | 6,612 | -42% | 1 | 1 | 0% | 1,679 | 1,528 | -9% | 0 | 0 | — |
case-17 | fail→pass | 5,580 | 2,015 | -64% | 1 | 1 | 0% | 923 | 787 | -15% | 0 | 0 | — |
case-18 | fail→fail | 5,370 | 6,309 | +17% | 1 | 1 | 0% | 750 | 814 | +9% | 0 | 0 | — |
case-19 | pass→pass | 8,950 | 4,469 | -50% | 1 | 1 | 0% | 1,413 | 1,173 | -17% | 0 | 0 | — |
case-20 | fail→pass | 3,017 | 2,622 | -13% | 1 | 1 | 0% | 385 | 889 | +131% | 0 | 0 | — |
case-21 | fail→pass | 10,457 | 6,496 | -38% | 1 | 1 | 0% | 1,691 | 1,455 | -14% | 0 | 0 | — |
case-22 | fail→fail | 5,951 | 2,044 | -66% | 1 | 1 | 0% | 760 | 680 | -11% | 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 19 counted toward the lift figure. The other 3 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 +32 percentage points is the difference between those two pass rates over the 19 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.