Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Diagnose wasted paid-ad spend, pacing, and allocation, then propose safe evidence-backed optimizations. Use for waste, negatives, budgets, bid changes, poor CPA or ROAS, underpacing, overspend, or scaling decisions.
.claude/skills/nowork-studio-paid-ads-optimize/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-01 | ✗→✓ | ▲ Improved | -51% | 0% |
| case-02 | ✗→✓ | ▲ Improved | -24% | 0% |
| case-03 | ✗→✓ | ▲ Improved | -18% | 0% |
| case-08 | ✗→✓ | ▲ Improved | 53% | 0% |
| case-13 | ✗→✓ | ▲ Improved | 9% | 0% |
Read ../shared/operating-contract.md and ../shared/measurement-framework.md. Review before changing anything.
Verify the conversion signal, period completeness, spend volume, attribution model, and recent account changes. Spend with no recorded conversion can indicate broken tracking or immature data; treat it as a hypothesis until the signal and volume support an intervention. Check landing-page or operational failures before blaming targeting.
Classify the bottleneck as query/audience quality, creative fatigue, delivery/rank, budget constraint, landing-page mismatch, tracking, or economics. Use the specialized Google, Meta, X, LinkedIn, Reddit, or TikTok skill for live diagnosis. For other platforms, analyze only the supplied or verified data.
Prefer this order: exclude an irrelevant query, placement, or audience; pause the narrowest losing unit; adjust budget or bid in a measured step; then consider structural change. For a reallocation, show the current and proposed allocations, the same total budget unless the user approves an increase, and the observable hypothesis.
Do not declare a loser from a few clicks. Set a threshold appropriate to the named target CPA, conversion lag, and channel role. Preserve upper-funnel and assisted-conversion context rather than judging all campaigns on last-click CPA alone.
Present each exact mutation with scope, current value, proposed value, currency exposure, rationale, and review date. After approval, execute only through the verified platform skill or connector, read back the result, and record the intervention's expected effect and guardrail. Revisit after the declared observation window instead of promising a generic ongoing watch.
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→pass | 46,362 | 21,816 | -53% | 1 | 1 | 0% | 7,530 | 3,692 | -51% | 0 | 0 | — |
case-02 | fail→pass | 23,736 | 18,572 | -22% | 1 | 1 | 0% | 3,900 | 2,981 | -24% | 0 | 0 | — |
case-03 | fail→pass | 30,212 | 27,302 | -10% | 1 | 1 | 0% | 4,906 | 4,036 | -18% | 0 | 0 | — |
case-04 | fail→fail | 14,970 | 12,480 | -17% | 1 | 1 | 0% | 2,018 | 2,182 | +8% | 0 | 0 | — |
case-05 | pass→pass | 10,292 | 9,813 | -5% | 1 | 1 | 0% | 1,755 | 2,060 | +17% | 0 | 0 | — |
case-06 | pass→pass | 98,761 | 37,413 | -62% | 1 | 1 | 0% | 8,219 | 8,275 | +1% | 0 | 0 | — |
case-07 | pass→pass | 15,472 | 34,859 | +125% | 1 | 1 | 0% | 2,349 | 2,437 | +4% | 0 | 0 | — |
case-08 | fail→pass | 14,662 | 21,048 | +44% | 1 | 1 | 0% | 2,039 | 3,113 | +53% | 0 | 0 | — |
case-09 | pass→pass | 17,937 | 18,790 | +5% | 1 | 1 | 0% | 2,651 | 2,860 | +8% | 0 | 0 | — |
case-10 | pass→pass | 12,585 | 10,829 | -14% | 1 | 1 | 0% | 1,719 | 1,898 | +10% | 0 | 0 | — |
case-11 | pass→pass | 18,556 | 23,396 | +26% | 1 | 1 | 0% | 2,767 | 2,672 | -3% | 0 | 0 | — |
case-12 | pass→pass | 15,063 | 14,827 | -2% | 1 | 1 | 0% | 2,179 | 2,281 | +5% | 0 | 0 | — |
case-13 | fail→pass | 10,239 | 10,826 | +6% | 1 | 1 | 0% | 1,703 | 1,857 | +9% | 0 | 0 | — |
case-14 | pass→pass | 30,169 | 12,951 | -57% | 1 | 1 | 0% | 2,316 | 2,064 | -11% | 0 | 0 | — |
case-15 | pass→pass | 11,226 | 9,689 | -14% | 1 | 1 | 0% | 1,529 | 1,521 | -1% | 0 | 0 | — |
case-16 | pass→pass | 14,439 | 12,051 | -17% | 1 | 1 | 0% | 2,176 | 2,223 | +2% | 0 | 0 | — |
case-17 | pass→pass | 15,188 | 15,363 | +1% | 1 | 1 | 0% | 2,022 | 2,674 | +32% | 0 | 0 | — |
case-18 | pass→pass | 9,641 | 94,362 | +879% | 1 | 1 | 0% | 1,197 | 1,532 | +28% | 0 | 0 | — |
case-19 | pass→pass | 29,139 | 15,167 | -48% | 1 | 1 | 0% | 2,179 | 2,339 | +7% | 0 | 0 | — |
case-20 | fail→fail | 26,880 | 7,530 | -72% | 1 | 1 | 0% | 2,142 | 1,411 | -34% | 0 | 0 | — |
case-21 | pass→pass | 15,523 | 15,661 | +1% | 1 | 1 | 0% | 2,203 | 2,587 | +17% | 0 | 0 | — |
case-22 | pass→pass | 21,195 | 22,002 | +4% | 1 | 1 | 0% | 2,842 | 2,317 | -18% | 0 | 0 | — |
case-23 | pass→fail | 9,159 | 4,331 | -53% | 1 | 1 | 0% | 1,390 | 766 | -45% | 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. 23 cases were attempted. The headline lift of +17 percentage points is the difference between those two pass rates over the 23 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.
| Model | Method | Date | Lift |
|---|---|---|---|
| gemini-3.6-flash | verified | 8/29/2026 | 0% |
Other measured skills in the registry, with their headline benchmark lift.