aiwf-avoid-ai-illustrations
AIWF generated-image quality guardrail for avoiding obvious AI illustration and image artifacts. Use when planning, prompting, reviewing, editing, or accepting generated images, illustrations, logos,
SKILL.mdskills/aiwf-avoid-ai-illustrations/SKILL.md
GitHubAIWF Avoid AI Illustrations
Overview
Use this skill to prevent generated images from looking broken, fake, or semantically wrong. It is an AIWF skill-pack lane for image-generation planning and final QA, not a broad UI/design cleanup skill.
Workflow
- Classify the surface before prompting or reviewing:
- freeform illustration or mood image
- photoreal person, face, hands, or product
- logo, icon, signage, or visual text
- diagram, flowchart, architecture figure, or process map
- chart, quantitative infographic, or data visualization
- Read
references/failure-modes.mdfor the relevant artifact categories. - Read
references/surface-playbooks.mdbefore generating, editing, or recommending an implementation route. - Read
references/review-checklist.mdbefore accepting a final asset. - Decide the exactness level:
- Flexible mood, scene, texture, and style can use normal image generation.
- Exact text, brand marks, charts, diagrams, anatomy-critical images, product claims, and instructional visuals need references, masks/control inputs, structured/vector rendering, code-backed charts, or explicit final inspection.
- Prefer correction over regeneration when the concept is good but a local defect exists: crop, inpaint, mask, replace text, redraw vector elements, or rerender charts from data.
Hard Rules
- Do not accept a generated chart as final unless values, axes, labels, title, and scale are checked against the source data.
- Do not accept generated logos or signage as final typography. Use generated output for concept direction, then render final text and marks as vector or editable design elements.
- Do not accept diagrams that need logic, arrows, labels, or editability as flat image guesses. Prefer Mermaid, Graphviz, SVG, draw.io, Figma, or code-backed layout.
- Do not assume high aesthetic quality means semantic correctness. Inspect prompt-critical entities, object counts, body parts, hands, text, labels, shadows, and object interactions.
- Do not over-prompt skin with extreme micro-detail when the user wants clean realism. Terms like "visible pores" can over-amplify blemish-like marks; prefer "natural, subtle skin texture" plus explicit absence of unwanted blemishes when that is the intent.
Output
For planning or prompt work, return the route, artifact risks, prompt/control guidance, and final QA gates.
For review work, return pass/fail findings by surface and state exactly what must be regenerated, inpainted, redrawn, or rendered deterministically.