Independent dimensioned review

When to use

Use this skill inside a governed Factory batch, after a round has artifacts and before the next round is planned. It produces one thing: a dimensioned advisory document for the round's evaluate step. It never generates images, never spends the allowance, and never delivers a verdict to the user.

Do not use it for direct (Baoyu) generation, for approvals or receipts, or as a second opinion on a batch that already has the user's decision. For presenting results and capturing the user's decision, hand off to the image-factory-judge skill and its conversation contract.

Workflow

1. Review in a separate context. Take only three inputs: the round's artifacts, the requirements the plan declared for each item, and the previous round's critique when one exists. Do not read the conversation that produced the images, and do not score artifacts you generated yourself in this session. The review is only as independent as its inputs.

2. Score named dimensions, not a mood. For each item, record bounded per-dimension scores, each with a short gap statement that names observable evidence ("the mark reads amber, not the declared deep emerald"), together with the total score between 0 and 1. Write the advisory file in the object form the evaluator accepts:

   {
     "item-01": {
       "score": 0.72,
       "reason": "close to the declared layout",
       "dimensions": [
         {"name": "character_identity", "score": 0.9, "evidence": "the left eyebrow scar and round glasses match the identity anchor"},
         {"name": "wardrobe", "score": 0.5, "evidence": "the coat is amber rather than the declared deep emerald"}
       ]
     }
   }

   A dimension whose statement names no observable evidence is left without evidence; the evaluator records it as incomplete, so it is reported but never counted as a gap. A missing dimension is honest; an invented one is not.

   For a series item, names are closed to character_identity, wardrobe, prop_continuity, style, scene_state, text_absence, and aspect_ratio. OCR, anatomy, identity similarity, and semantic continuity remain advisory observations; do not rewrite them as deterministic failures.

3. Let the score fall. Compare against the previous critique for consistency, but do not match or raise the previous total to appear consistent. When this round is worse, record the lower total; the drop is kept and reported as a regression. Do not settle for a flattering number.
4. Stay inside the platform. Every gap you record must be addressable by a different prompt or different reference images. Never record a gap whose fix requires a size, quality, background, image count, or model change, and never describe a further attempt that would start without an explicit instruction for that item.
5. Stop at the document. Hand the advisory file to the evaluate step of image-factory-judge. The deterministic gates, the human labels, and the user conversation stay there. This skill is part of the evidence chain, not of the conversation.

Never do

• Never score artifacts you generated yourself, and never read the generation conversation as review input.
• Never present the critique, to the user or in the record, as a verdict. It is advisory: it cannot fail a batch and cannot outrank a human rejection.
• Never raise a total to keep continuity with a previous round when the images got worse.
• Never record a gap that requires a generation parameter this plugin does not control, and never describe an attempt that would occur without an explicit per-item instruction.
• Never copy the body of another skill; hand off by name.
• Never promise a visual match beyond what the recorded evidence shows.