creative-qa（创意质量保证）

Creative QA

Follow shared release-shell rules in:

• postplus-shared release-shell rules

Use this skill after a human has actually reviewed an asset.

This skill is for:

• recording human judgments about generated assets
• preserving why something was approved, revised, or rejected
• pointing feedback back to the stage that should change
• turning review notes into structured data for later analysis

This skill is not for autonomous AI approval.

Core Idea

The first version should assume:

• humans decide quality
• AI may help summarize or prefill observations
• only human-confirmed feedback becomes the durable QA record

If there is no human feedback yet, there may be no QA record yet. That is acceptable.

Human Rule

Do not invent a final verdict on behalf of a human reviewer.

Allowed:

• turn human notes into structured fields
• normalize categories
• suggest likely blame stages

Not allowed:

• silently approving an asset
• replacing a human verdict with an AI guess

Objects

1. QA Report

One review record tied to one asset version.

Should include:

• qaReportId
• targetObjectType
• targetObjectId
• targetVersion
• reviewer
• verdict
• goodReasons
• badReasons
• issueCategories
• blameStage
• proposedAction

2. Feedback Record

An optional follow-up object that can be consumed by rerun workflows.

Should include:

• feedbackId
• qaReportId
• feedbackCategory
• feedbackText
• dependencyImpact
• rerunTarget

Scope

This skill should support human review records for:

• image
• voice_take
• video_render
• final_export

Local Persistence Convention

Store QA next to the asset being reviewed.

In the released product shell:

• keep draft QA notes or intermediate review payloads under <work-folder>/.postplus/creative-qa/
• keep the final confirmed QA record next to the reviewed asset or final deliverable

One possible project-local layout is:

<work-folder>/.postplus/creative-qa/video-render/
  qa-v1.json

or:

reviews/voice-take-1.review.json

The key requirement is stable linkage back to the reviewed object.

Review Categories

Common issue categories:

• lip_sync
• persona_drift
• audio_style
• audio_pacing
• hook_weak
• ad_like
• ugc_native_feel
• visual_realism
• subtitle_accuracy
• mixed

Common blame stages:

• image
• script
• voice
• render
• subtitle
• mixed

Output Rule

The QA layer should answer:

• what was reviewed
• who reviewed it
• what they decided
• why they decided it
• what should happen next

Read references/qa-schema.md when creating or updating these files.