AI Image Generation
Purpose
This skill produces AI-generated images that fit a stated purpose, audience, and placement. It covers defining the brief, writing the prompt, generating with an available image tool, reviewing the result for defects, and delivering a correctly sized, correctly licensed file with alt text.
When to Use
- A task asks for an AI-generated image, illustration, hero image, or artwork.
- A blog post, landing page, or presentation needs an original visual and no suitable stock photo exists.
- A concept needs a quick visual mockup before a designer produces the final art.
- A series of images must share one consistent style across articles or pages.
When Not to Use
- The task needs a real photograph of a real place, person, or product — use a camera or a licensed stock photo instead; AI output is not evidence.
- The task needs a precise diagram, chart, or data visualization — use
infographic-generationor a charting tool instead. - The task needs only a prompt, not a generated file — use
image-prompt-engineeringinstead. - The task needs a social preview image — use
og-image-generationortwitter-card-imagesinstead, which cover the exact dimensions and tags.
Prerequisites
- A clear brief: subject, purpose, placement, audience, and mood.
- Access to an image-generation tool in the current environment, or a named external tool the user can run.
- The target dimensions or aspect ratio for the placement.
- Any brand constraints: palette, style, logo usage, and forbidden content.
- Knowledge of where the file will be stored and what filename to use.
Official Documentation
- OpenAI platform docs — image model guides and usage policies for OpenAI image tools.
- Midjourney documentation — Midjourney parameters, styles, and community guidelines.
- Google AI for Developers — Gemini image generation guides and safety documentation.
- Unsplash — a stock-photo alternative when a real photograph is the better choice.
- Compatibility: image-model names, capabilities, and usage terms change frequently. Do not name a specific model version from memory — check the linked docs for the current model before promising a capability, and record the exact tool and model version used with each delivered image.
Required Workflow
- Confirm the brief in writing: subject, setting, style, mood, aspect ratio, dimensions, and where the image will appear.
- Check the brand constraints and any content restrictions before prompting.
- Write the prompt with
image-prompt-engineering: subject, setting, style, lighting, composition, color palette, and what to avoid. - Generate the image with an available image tool. If no image-generation tool is available in this environment, stop and hand the user the finished prompt plus the target dimensions instead of claiming an image was produced.
- Review the output at full size: hands, faces, text, logos, anatomy, perspective, and any garbled details. Regenerate or revise the prompt when a defect is visible.
- Check the image against the brief: correct subject, correct mood, correct orientation, and no unwanted elements or text.
- Resize or crop to the target dimensions and save in the right format
(see
image-format-conversionandimage-compression). - Write alt text for the final image (see
image-accessibility). - Record the prompt, tool, settings, and license or usage terms alongside the delivered file so the image can be regenerated or audited later.
Implementation Standards
- One image per brief. A new subject, style, or placement gets a new prompt and a new review; do not reuse a prompt blindly across placements.
- Generate at or above the final display size, then downscale. Never upscale a small generated image and present it as high quality.
- Keep the prompt, tool name, model or version if known, seed or settings if the tool exposes them, and the generation date with the deliverable.
- Filenames are lowercase, hyphenated, and descriptive, for example
sourdough-starter-glass-jar-hero.jpg. - Never present an AI-generated image as a real photograph. Label it as AI-generated wherever the audience could mistake it for a real photo, and follow the disclosure rules of the platform where it is published.
- Text inside generated images is unreliable. Keep in-image text to a few short words at most, and verify every character in the output.
Best Practices
- Generate several variations and pick the strongest one rather than accepting the first result.
- Describe the composition explicitly: subject placement, background simplicity, and empty space for headlines or overlays.
- Use a consistent style phrase across a series so related images look like they belong together.
- Prefer simple backgrounds and one clear focal point; busy scenes hide defects and distract from the message.
- Keep people out of the image when the brief does not need them, since faces and hands are the most common source of visible defects.
- Save the best prompt for reuse, and note which variations failed and why.
Security Considerations
- Never prompt for a real, identifiable person, and never generate imagery that impersonates someone or reproduces a private individual's likeness.
- Do not prompt for copyrighted characters, logos, watermarks, or distinctive artwork in a way that reproduces someone else's protected work.
- Do not include personal data, private photos, API keys, or internal documents in a prompt sent to a third-party image service.
- Check the image service's terms for ownership and commercial-use rights before using an image in published or commercial work.
- Review output for accidental inclusion of readable text such as signs, screens, or documents that could leak information.
Performance Considerations
- Generate at the smallest size that still meets the placement's display size; oversized generations cost more time and storage.
- Compress the delivered file (see
image-compression) so page weight stays low; a hero image should not dominate the page's total size. - Prefer modern formats such as WebP or AVIF for web delivery when the platform supports them, with a JPEG fallback only where needed.
- Reuse one generated image across placements only when the crop and dimensions still suit each placement; otherwise generate or crop per placement.
Common Mistakes
- Accepting the first generation — the first result often has subtle defects; always generate variations and review at full size.
- Vague prompts such as "a nice picture of bread" — a vague prompt gives a generic result; specify subject, style, lighting, and composition.
- Claiming an image was generated when no image tool was available — if there is no tool, deliver the prompt and dimensions, not a fake file.
- Using AI output as documentary evidence — generated images are illustrations, not proof of a real event, place, or product.
- Ignoring licensing — check the tool's terms before commercial use; not every image service grants the same rights.
- Letting the model render long text — generated text is frequently misspelled; add real text with an editor or keep it to a few words.
Examples
Example brief, written before any prompt:
Subject: a glass jar of bubbling sourdough starter on a wooden counter
Setting: bright home kitchen, morning light, blurred background
Style: warm editorial food photography look, shallow depth of field
Mood: inviting, calm, homemade
Aspect ratio: 16:9, landscape, for a blog hero image
Avoid: people, hands, text, logos, cluttered counters
Example prompt built from that brief:
Editorial food photograph of a glass jar of bubbling sourdough starter
on a light wooden counter, bright morning kitchen light, soft blurred
background, warm tones, shallow depth of field, 16:9 landscape
composition with empty space on the left for a headline. No people,
no hands, no text, no logos.
Example delivery note recorded with the file:
File: sourdough-starter-glass-jar-hero.jpg
Tool: <image tool used> | Prompt: see prompt above
Size: 1600x900 | Format: JPEG, quality 82
Alt text: Glass jar of bubbling sourdough starter on a wooden counter
in morning light.
License: per the image tool's terms; verify before commercial reuse.
Validation Checklist
- The brief (subject, style, dimensions, placement) was confirmed before generation.
- The image was generated with a real, available image tool, or the prompt and dimensions were handed over instead.
- The output was reviewed at full size and has no visible defects in faces, hands, text, or anatomy.
- The image matches the brief's subject, mood, and orientation.
- The file is the correct size and format for its placement.
- Alt text is written for the final image.
- The prompt, tool, and usage terms are recorded with the file.
- The image is labeled AI-generated wherever it could be mistaken for a real photograph.
Expected Deliverables
- The final image file at the target dimensions, in the agreed format.
- The prompt used, plus the tool name and any settings worth recording.
- Alt text for the final image.
- A short note on usage terms and any disclosure the publisher must add.
- The list of rejected variations, if any, with the reason each was rejected.
References
- OpenAI platform docs — image generation guides and usage policies.
- Midjourney documentation — parameters and style guidance.
- Google AI for Developers — image generation guides and safety documentation.
- Related skills in this repository:
image-prompt-engineering,blog-featured-images,image-compression,image-format-conversion,image-accessibility,brand-visual-consistency.