paid-ads-optimize
GitHub诊断付费广告支出浪费、投放节奏及分配问题，提供基于证据的安全优化建议。涵盖预算调整、出价变更、CPA/ROAS分析及扩量决策，强调先验证数据再执行操作。
Trigger Scenarios
Install
npx skills add nowork-studio/notfair-plugin --skill paid-ads-optimize -g -y
SKILL.md
Frontmatter
{
"name": "paid-ads-optimize",
"description": "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.",
"argument-hint": "<campaign, target CPA\/ROAS, or 'find waste'>"
}
Paid Ads Optimization
Read ../shared/operating-contract.md and ../shared/measurement-framework.md. Review before changing anything.
Diagnose before cutting
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, or LinkedIn skill for live diagnosis. For other platforms, analyze only the supplied or verified data.
Rank reversible moves
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.
Approval and follow-up
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.
Version History
- 8d9d54b Current 2026-08-19 23:11
