Image generator workflow
Make an image that belongs to its destination, rather than an attractive generic illustration. A generation is complete only when the selected image has been inspected at full size and at its intended crop. This skill can use any available image generation or editing tool; do not claim that installing the skill adds a renderer or guarantees a particular model.
Inputs: define the image job
Read the article, product page, or brief before drafting a prompt. Record: the single idea the image must communicate; where it will appear; audience; target dimensions, aspect ratio, and safe crop; brand palette and typography; required and forbidden subjects; and which details must be factually exact. Distinguish a decorative hero, a product screenshot, a diagram, and a social preview: each has a different accuracy and legibility burden.
Inspect supplied photos, logos, screenshots, and visual references. Name what each reference contributes—composition, lighting, material, mood, or an exact object. Do not copy an unrelated reference's subject just because its style is appealing. When a reference is an existing image to modify, use the tool's edit or reference mechanism, and state what must remain unchanged. Keep track of source and usage rights for publishable work.
Workflow: choose the method and tool
- Generate when the subject can be invented and no pixel-accurate source exists.
- Edit when the supplied image already has the right scene, subject, or composition and only a bounded change is needed.
- Composite or draw verified assets for exact logos, UI, quoted text, numbers, charts, or diagrams. Do not rely on generated pixels to reproduce these exactly.
- Use existing photography or a screenshot when authenticity is the main claim. Generating a fictitious product interface or event photo can mislead readers.
Check the selected tool's current capabilities and limits in its official documentation. For OpenAI tools, distinguish single-image generation or edits from a multi-turn image workflow. If the user asks for a particular model, verify that the tool actually offers it; if model selection is hidden, report that plainly. Prefer the best available model for the task within the user's cost and time constraints, not an unverified model name.
If no renderer is available, return a production-ready brief and prompt rather than pretending an image was created. Do not buy credits, subscribe, or publish the image without authorization.
3. Write a testable art direction
Specify subject and action, camera or viewpoint, framing, lighting, palette, material, texture, background, and intended emotional tone. Add constraints that matter to the destination: open space for a headline, focal point away from the crop edge, mobile legibility, or no embedded text. Avoid piles of style adjectives that compete with the subject. Create a small number of meaningfully different candidates by changing a clear design choice—viewpoint, visual metaphor, or setting—rather than adding random variations.
A prompt structure that travels across tools:
Create [asset type] for [placement and audience]. Show [specific subject doing a visible action] in [setting]. Use [camera/framing], [lighting/material], and [palette]. The focal point is [location]; keep [area] clear for layout. Use the supplied [reference] for [specific property only]. Preserve [facts or supplied elements]. Avoid [two or three concrete failure modes]. Output [aspect ratio/size]. No generated text or invented logo.
Worked example: For a Thrive article about using references in image generation, show a designer's hands comparing three distinctly different photographic prints on a warm light table. Give each print a coherent subject and realistic lighting; leave quiet space along the upper edge for the article title. Use a supplied desk photograph only for lighting and camera angle. Do not put Thrive's logo, interface, or headline inside the generated pixels; those belong in the final layout as verified assets.
4. Select, refine, and inspect
Compare candidates against the brief, not against each other alone. Reject a beautiful image that fails the topic or implies a false product capability. Choose one candidate and make one targeted edit at a time where possible—e.g., “preserve the hands and print arrangement; reduce the distracting glare on the central print.” Reinspect the whole image after every edit; a local fix can damage another area.
Check at 100% and at the actual placement width:
- Subject and action match the article or product claim.
- Anatomy, reflections, perspective, repeated objects, and object boundaries hold up.
- Text, icons, UI, diagrams, and logos are either verified or absent.
- Important details survive desktop, mobile, and social crops; the focal point is not hidden by overlays.
- Contrast and visual complexity leave room for any real text added later.
- Color and texture look consistent with the site, without uncanny polish or generic AI motifs.
If the image contains a real person, product, place, or event, verify that the result does not suggest documentary truth where none exists. For charts and diagrams, validate each label and value independently; redraw when exactness matters.
5. Export and hand off
Export the chosen master and placement-ready variant in a suitable web format and dimensions. Preserve the source or edit provenance and any license information needed for publication. Provide descriptive alt text that explains the image's useful content rather than repeating the article title. State the intended crop, file location, tool used, edits made, and any accuracy limit. If the image was not rendered, label the output brief only.
The image generator skill guide gives an editorial use case; the current OpenAI image prompting guide has tool-specific prompting patterns.
