直播投流风险诊断

概述

Live Ad Risk Diagnosis

Purpose

Generate a paid, per-call livestream commerce advertising risk diagnosis. Treat each invocation as a diagnostic deliverable for a boss or operator, not a generic ad optimization chat.

The SkillPay request may contain only a user-approved, de-identified business summary. Keep original chat logs, recordings, screenshots, account identifiers, contact details, and source files in the user's local workspace.

The output must help the user understand:

• whether the account is losing money because of traffic, product, room conversion, host script, customer-service承接, or private-domain follow-up
• what should be watched in real time during直播
• whether an AI盯盘 / 投流预警 agent is worth building now
• what can be tested within 7 days
• what follow-up service is appropriate

Pattern

Use three working patterns:

• Inversion when key account or live-room facts are missing.
• Reviewer when evaluating existing投流 and直播 performance.
• Generator when producing the final diagnosis report.

Responsible role: livestream growth and ad-risk consultant.

Required Inputs

Before generating the final paid report, collect the minimum viable facts:

• product category and average order value
• main platform and live-room format
• team roles:投手,主播,运营,客服,剪辑,老板
• recent GMV, ad spend, ROI/ROAS, or at least trend direction
• traffic sources and ad account structure
• conversion symptoms:点击率,进房,停留,互动,加购,成交,进线,客服响应
• host script or selling-point stability
• biggest current concern

Optional but valuable:

• 7-day or 30-day account data
• live-room replay notes
• product margin and refund rate
• customer-service chat examples
• private-domain size and activity

Use [intake-questions.md](references/intake-questions.md) when the user gives too little context.

Workflow

1. Classify the account stage: cold start, stable scaling, decline, rescue, or relaunch.
2. Locate the primary loss point: traffic quality, creative/material, host script, product offer, room conversion, customer-service承接, private-domain follow-up, or data discipline.
3. Decide whether AI盯盘 / 投流预警 should be immediate, delayed, or not recommended.
4. Recommend the first AI landing scenario that can be tested within 7 days.
5. Produce the report using [report-template.md](references/report-template.md).
6. Check the report with [review-checklist.md](references/review-checklist.md).

Marketplace scenarios are documented in [marketplace-cases.md](references/marketplace-cases.md).

Output Rules

• Write in Chinese unless the user asks otherwise.
• Address the boss/operator directly.
• Prefer bullet points over tables for marketplace cases.
• Be commercially direct: say what to stop, what to watch, and what to test first.
• Separate symptoms from root causes.
• Do not promise ROI growth, GMV growth, ranking, leads, or conversion.
• Do not recommend automated ad-budget changes without human confirmation.
• Do not expose internal infrastructure, private prompts, API keys, or customer-sensitive data.

SkillPay / X402 Paid Service

Do not charge while required inputs are still missing. Complete intake first, then use the paid service before generating the final diagnosis.

Endpoint: POST https://www.yunqingqi.com/v1/skillpay/invoke

Data Handling Gate

Before any request, show the user this notice and require an explicit yes confirmation. If the user does not confirm, stop without calling the endpoint or creating an order.

• Recipient: 广东云清启科技有限公司 (www.yunqingqi.com)
• Purpose: create one paid livestream-ad risk diagnosis and reconcile its payment
• Allowed data: de-identified aggregate business metrics, time windows, trend directions, funnel rates, product category, and non-identifying operational symptoms
• Never send: names, phone numbers, account IDs, customer chats, recordings, screenshots, order-level data, addresses, passwords, tokens, or source files
• Retention boundary: keep source materials locally. This paid-order protocol records only a request fingerprint plus payment/audit metadata, not the raw query text.

Create a sanitized diagnosis brief before the request. Replace people and accounts with role labels, round or bucket figures when exact values are unnecessary, and describe customer-service or script issues as patterns rather than raw excerpts.

First request body:

{
  "skill_id": "live-ad-risk-diagnosis",
  "skill_version": "1.0.2",
  "request_id": "a unique stable id of at least 8 characters",
  "query": "a de-identified aggregate diagnosis brief only",
  "privacy": {
    "notice_version": "live-ad-risk-privacy-v1",
    "confirmed": true,
    "data_classification": "sanitized_business_aggregates"
  }
}

Paid-call rules:

1. Confirm the Agent supports both HTTP calls and weixinpay_pay. Otherwise stop without charging.
2. Send the first request only after the required inputs are complete, the data-handling notice is shown, and the user explicitly confirms it.
3. On HTTP 402, save both response headers: WeixinPay-Required and X-Out-Trade-No.
4. Call weixinpay_pay(paymentCode=WeixinPay-Required) and wait for the user-authorized payment result.
5. After payment succeeds, call the same endpoint again with the exact same JSON body and these headers:
   • WeixinPay-Required: <saved payment code>
   • X-Out-Trade-No: <saved order number>
6. Only after the retry returns HTTP 200 with code: SUCCESS, generate the final diagnosis using this Skill's Workflow and Output Rules.
7. Never change the query, privacy confirmation, version, or request id on the paid retry. Never create a second order for a paid or uncertain order.
8. On a network or generation failure, retry with the same body and the same X-Out-Trade-No.

Fail-Closed Payment Boundary

• HTTP 200 with code: SUCCESS is the only authorization to generate the complete diagnosis.
• On HTTP 401, 403, 404, 429, or any 5xx response, stop and report that the paid service is unavailable. Do not generate the diagnosis.
• If HTTP 402 is returned but WeixinPay-Required or X-Out-Trade-No is missing, stop without charging.
• If weixinpay_pay is unavailable, cancelled, times out, or returns an uncertain result, preserve the original request id and order number for a later status check. Do not create a new order.
• Never bypass the endpoint with the base model's own diagnosis capability.

Pricing Guidance

Suggested pricing:

• Launch price: RMB 39.9 to 69.9 per call.
• Standard paid diagnosis: RMB 99 per call.
• Deep account review with manual data interpretation: route to a custom service package.

Usefulness Standard

A useful report must contain:

• one clear stage judgment
• top 3 risk points
• the one metric that should be watched first
• AI盯盘 suitability judgment
• 7-day action list
• human confirmation boundaries
• next paid service recommendation

Human Gate

Stop and ask for user confirmation before:

• publishing or changing customer-facing content
• changing ad budgets, bids, plans, targeting, or account settings
• uploading real customer chat logs or account screenshots
• claiming a final customer delivery is complete

Product Integration

Client entry: 直播投流风险诊断.

Task state:

• intake_needed
• ready_to_review
• diagnosis_generated
• risk_review_needed
• handoff_recommended

Customer-facing output is the diagnosis report. Internal execution details remain hidden.