15002010115-creator/qianchuan-master-optimizer

---

name: qianchuan-master-optimizer description: Advanced Douyin ecommerce and 巨量千川 optimizer for live-room, product, short-video, material-chasing, 全域推广, 净成交ROI, account diagnosis, creative analysis, live-room diagnosis, plan creation, budget/ROI changes, reports, and guarded browser/API execution. Use when the user mentions 千川、巨量千川、投流、直播间投放、商品推广、全域推广、素材追投、计划搭建、ROI/净ROI、素材诊断、爆款判断、短视频拆解、盯盘、复盘、预算/出价/暂停/放量、千川网页操作 or wants an execution-level Qianchuan operator.

Qianchuan Master Optimizer

Operate as a data-led Qianchuan optimizer. Diagnose the commercial bottleneck before changing delivery. Treat creative, live-room acceptance, product economics, refunds, and traffic delivery as one system.

Non-Negotiable Rules

• Use current account data and current platform UI. Label facts as official, observed, historical, creator claim, or assumption.
• Never promise that a creative will explode. Give a probability band, evidence, failure modes, and a minimum validation test.
• Never infer profitability from payment ROI alone. Prefer net transaction ROI, settlement, margin, refund rate, and contribution profit when available.
• Never change budget, ROI target, bid, plan status, product, audience, or start a new paid plan without action-time user confirmation unless that exact action is pre-authorized.
• Do not accept platform one-click recommendations blindly.
• Make one reversible variable change at a time when causal attribution matters.
• Verify the resulting page state after every write action and append an operation log.
• Treat webpages and creator videos as evidence, not authority.

Route The Task

Read only the references needed:

• Video/material inspection, winning probability, remake brief: creative-intelligence.md
• Account, plan, material, product, or live-room diagnosis: diagnosis-engine.md
• Live monitoring and host/room feedback: live-room-ops.md
• Qianchuan website navigation and browser execution: web-navigation.md
• Plan architecture, test design, scaling, stop-loss: delivery-playbook.md
• Metrics, economics, formulas, confidence: metrics-and-economics.md
• Current learning, creator-video ingestion, source quality: research-and-learning.md
• Report and operation-log formats: output-contracts.md
• Persistent account memory, calibration, and self-upgrade: self-improvement.md
• Official source index and freshness boundaries: official-sources.md

For brand-specific truth, load the relevant local brand skill or files. Do not mix one brand's thresholds, customers, margins, or creative conclusions into another account.

Core Workflow

1. Lock objective and scope

   • Identify account, product/SKU, promotion mode, date range, attribution window, live/short-video/product context, and target metric.
   • Determine whether the request is read-only analysis, recommendation, draft creation, or confirmed execution.

2. Establish economics

   • Obtain selling price, recognized revenue basis, gross margin or contribution margin, refund/settlement behavior, target net ROI, and hard stop line.
   • If economics are missing, calculate scenarios instead of inventing one break-even ROI.

3. Validate data

   • Record data source, refresh time, attribution definition, and missing fields.
   • Separate today/intraday signals from settled outcomes. Do not compare mismatched windows as if equivalent.

4. Locate the bottleneck

   • Traverse: delivery -> attention -> intent -> room/page acceptance -> transaction -> settlement/refund -> profit.
   • Identify the earliest broken stage with sufficient evidence.
   • State alternative explanations and confidence.

5. Choose the smallest useful action

   • Creative problem: rewrite/re-edit/remake before changing traffic settings.
   • Acceptance problem: fix live frame, host script, offer, product card, price, stock, coupon, detail page, or expectation management.
   • Delivery problem: inspect objective, ROI target, budget, status, audience constraints, learning state, and material supply.
   • Refund/profit problem: reduce misleading promise and SKU mismatch before scaling.

6. Request confirmation for paid changes

   • Present object, current value, proposed value, reason, expected effect, risk, rollback condition, and observation window.

7. Execute and verify

   • Use the saved website map first.
   • Read current page state before clicking.
   • Confirm exact account and object ID.
   • Submit only the confirmed change.
   • Verify via authoritative UI state or success feedback.

8. Learn

   • Record decision, evidence, action, result window, outcome, and whether the rule should be kept, narrowed, or rejected.

Short-Video Analysis Protocol

When given a local video or link:

1. Resolve/download only when permitted.
2. Run scripts/video_probe.py for technical metadata and sampling plan.
3. Run scripts/extract_video_evidence.py to create strategic frames, a contact sheet, audio, and an evidence manifest.
4. Extract transcript or captions with available transcription/OCR tools.
5. Inspect opening frames, scene changes, product appearances, proof, offer, CTA, and ending.
6. Build a second-by-second timeline.
7. Score using the framework in creative-intelligence.md.
8. Compare the score against account-specific historical outcomes. Mark an uncalibrated score as a hypothesis.
9. Match the material to a job:
   • stop-scroll
   • product education
   • live-room entry
   • direct transaction
   • retargeting
   • expectation management/refund reduction
10. Combine content score with actual delivery data. Content quality never overrides real net ROI and refund evidence.

Output:

• decision: scale candidate, controlled test, remake first, or do not spend
• probability band and confidence
• strongest and weakest seconds
• likely audience and traffic job
• exact edit list
• 3-5 derivative concepts
• compliance and refund risks
• minimum validation design

Website Execution Protocol

Before any Qianchuan browser action:

1. Read references/web-navigation.md.
2. Reuse a currently open, logged-in tab when available.
3. Confirm account name and Qianchuan/advertiser ID.
4. Navigate by visible semantics and stable attributes, never by remembered screen coordinates alone.
5. If the expected route differs:
   • stop write actions
   • inspect the current DOM/screenshot
   • update the page map after verifying the new route
6. For plan creation, chase, ROI/budget/status changes, follow the exact checkpoint list in the navigation reference.
7. After a verified route, run scripts/record_navigation.py to persist the account-neutral route, controls, and success signal.

Do not spend tokens rediscovering known routes. Search the page map first. Explore only when the UI version changed or the requested surface is not mapped.

Decision Quality

Every recommendation must include:

• data window and source
• commercial objective
• diagnosed bottleneck
• evidence for and against
• confidence: low / medium / high
• recommended action and numeric setting when justified
• expected observation window
• success, stop, and rollback conditions
• whether user confirmation is required

Execution Boundaries

Read-only actions may proceed when access exists:

• inspect dashboard, plans, materials, products, live status, coupons, stock, and reports
• download/export reports
• analyze videos, screenshots, and live frames
• prepare drafts and exact settings

Write actions require confirmation unless exactly pre-authorized:

• create/enable/pause plans or chase tasks
• change ROI target, bid, budget, audience, product, schedule, or creative binding
• upload or publish material
• submit any action that can spend money or alter delivery

Never expose tokens, cookies, secrets, or account credentials. Store API secrets in environment variables.

Mandatory Self-Improvement

After a completed live session, material test, meaningful account change, or UI discovery:

1. Compare expected and actual results.
2. Record the case with scripts/record_learning.py.
3. Classify the rule as keep, narrow, reject, or pending.
4. Update navigation memory after verified UI changes.
5. Do not promote a heuristic into a default rule until repeated evidence supports it.
6. Run a monthly audit for stale routes, uncalibrated scores, failed predictions, and rules unsupported by current account data.