--- name: oneirloom-result-diagnosis description: Compare an actual or user-reported result miss with the source, current intent, and submitted prompt. Separate source-to-prompt from prompt-to-result errors and return a complete correction that preserves successful relations. Ordinary requested edits need no diagnosis. --- # Result diagnosis Read the [shared interaction contract](../oneirloom/references/interaction-contract.md) before the first substantive response unless its unchanged content is already available. For person prompts, including new images and grid panels, read [person defaults and proportions](../oneirloom/references/person-prompts.md). Use this capability for an actual or user-reported result miss, or a persistent failure after focused corrections. A request to change a background or color without a reported miss is an ordinary revision under the shared contract, not an automatic diagnosis. Distinguish the model and entry that produced the sample from the desired target. ## Compare the available evidence For graphic results, read [graphic review](../oneirloom-style-design/references/graphic-review.md) directly. Do not load the graphic creation entry unless findings require renewed understanding or design. Photography and other tasks use the same attribution rules below. 1. Record the requested change, confirmed target, observed or reported deviation, successful parts, main correction dimension, and acceptance basis. With no visible result, use the report conditionally and do not claim inspection. 2. For reconstruction, read [visual analysis](../oneirloom-visual-analysis/SKILL.md) and its reconstruction workflow. Compare the source, complete submitted prompt when available, and result against the same structural anchors: held-object direction and grip, body-relative covering span, and garment coverage, openings, connectors, and layers. Check hosiery separately from footwear. Distinguish structural misses from taste; hidden details remain unknown. 3. Compare source-to-prompt fidelity separately from prompt-to-result compliance. An omitted or semantically wrong relation is different from a clearly stated requirement the output ignored; both can occur in one result. Covering a skirt beside the hip differs from concealing the thigh-root region. A stated instep opening that appears filled is an output deviation, not a missing instruction. 4. Attribute causes only as far as evidence supports. Without the actual prompt, analysis, integration, and generation attribution remain unresolved. Missing execution metadata leaves entry, reference-channel, crop, enhancement, and randomness explanations uncertain. A visible mismatch does not locate the failed stage or prove which skill revision was exercised. Do not infer a model rule from one sample. 5. Correct the requested main dimension where possible. A failed result is evidence, not a replacement target. Preserve successful source- or user-grounded relations and compare the full revision with confirmed scale, crop, occlusion, and prominence. Explain necessary coupled changes. For a local image correction, prefer editing with confirmed masks or reference channels when appropriate instead of repeatedly regenerating everything. 6. For reconstruction corrections, run the structural counterexample check on the complete revision in every delivered language. Naming a prop or garment category alone is insufficient. If local fullness has spread to the whole body, revise globally scoped size terms while preserving requested local volume, other features, camera, and framing. Do not layer negative body-size commands over the cause. ## Bound a persistent correction If the same critical miss remains after two controlled corrections, inspect the actual model, reference channel, prompt enhancement, aspect ratio, and available controls. Do not keep adding synonymous adjectives. Missing access blocks only the unsupported explanation or operation. For repeated crop misses, check whether aspect ratio, subject occupancy, pose, and other requirements conflict. When exact boundaries matter, use supported cropping, editing, or reference controls; a suitable post-generation crop is also possible. Report post-processing separately from direct model compliance. ## Correct a real product Read [product art direction](../oneirloom-product-art-direction/SKILL.md) for any product miss. Compare original evidence and the current agreed concept with the submitted prompt and result separately. A miniature world may be intentional; attractive styling does not justify invented filling, altered packaging, or a wrong count. Diagnose against the user's replacement direction when one exists, retaining valid product facts. Describe the supported package geometry and target relations affirmatively. Preserve successful background, style, and copy unless correction requires explained reflow. Never adopt generated details as real-product evidence. The product method owns unresolved labels and exact-text dependencies. Label user reports and synthetic descriptions accurately, and leave visual inspection and unsupported execution causes unresolved. ## Deliver the correction Return the complete revised prompt or requested artifact under the shared interaction contract, with actual review status. For prompt delivery, apply its final-assembly and fenced-block rules. For a conceptual graphic failure, read [graphic design](../oneirloom-style-design/SKILL.md) and repair the responsible early stage. With the design intact, correct local errors in available editable source or an appropriate image edit. Change production path only when the deliverable or observed limits justify it. Keep diagnosis, counterexamples, and acceptance lists outside copyable generation prose. Preserve matched source relations and leave hidden construction unknown. A compact internal note can record the target, deviation, changed dimension, preserved relations, and decisive acceptance criterion. It is not another ledger.