creative-qa

Creative QA

Use When

• Record human-in-the-loop quality judgments for generated images, voice takes, video renders, and final exports.
• Use after a person has reviewed an asset and you need structured verdicts, reasons, issue categories, blame stage, proposed action, and rerun guidance.

Do Not Use When

• No human verdict exists yet.
• The user wants autonomous AI approval.
• The task is prompt critique before generation; use preflight or prompt QA instead.

Required Input

• Human reviewer verdict and reasons.
• Target object type, id, and version.
• Reviewer and review timestamp.
• Proposed action.

Human Rule

Humans decide quality. AI may summarize notes, normalize categories, and suggest likely blame stages, but only human-confirmed feedback becomes the durable QA record.

Do not invent approval, rejection, or reviewer intent.

Default Workflow

1. Confirm the reviewed asset: type, id, version, and optional campaign/persona/concept linkage.
2. Confirm the human reviewer, timestamp, verdict, and reasons.
3. Normalize reasons into goodReasons, badReasons, issue categories, and blame stage.
4. Decide proposed action: approve, rerun a specific stage, revise source input, or reject.
5. If rerun guidance exists, build a feedback handoff with a concrete rerunTarget.
6. Store or return the structured QA record next to the reviewed object or in the work folder's QA area.

Supported Objects

• image
• voice_take
• video_render
• final_export

Verdicts

Valid verdicts are approved, revise, and reject.

For revise or reject, require non-empty badReasons and issueCategories. For approved, these may be empty arrays.

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

Use blame stage as diagnostic guidance, not as a substitute for human verdict.

Stop Conditions

• Stop when required user intent, source evidence, or owned input artifacts are missing and guessing would change the result.
• If an owned CLI or script command fails, report the exact error and stop. Do not bypass the failure with metadata-only answers, readiness probing, local payload rewrites, fallback providers, or unpublished tools.

Output Shape

qaReport:

• qaReportId, targetObjectType, targetObjectId, targetVersion, campaignId, personaId, conceptId, reviewer, reviewedAt, verdict, goodReasons, badReasons, issueCategories, blameStage, scores, proposedAction, status

Optional feedbackHandoff:

• feedbackId, qaReportId, targetObjectType, targetObjectId, feedbackCategory, feedbackText, dependencyImpact, rerunTarget

Anti-Patterns

• silently approving an asset
• replacing a human verdict with an AI guess
• using vague stage labels as rerun targets
• losing linkage to the reviewed object version
• preserving subjective chat notes without structured reasons

Handoff

• Return the structured QA record structured output or an explicit blocker. If feedback is present, hand off to the named rerun target with the QA record attached.

Public Command Boundary

• Choose the smallest matching command or workflow from the user input and run it directly.
• If an owned CLI or script command fails, report the exact error and stop. Do not bypass the failure with metadata-only answers, readiness probing, local payload rewrites, fallback providers, or unpublished tools.