Prompt Craft: A Step-by-Step Method for Better AI Images

This is an independent educational article on jsrkalonda.com, not vendor documentation.

Most people write image prompts the way they might guess a password: type something, hope it works, and if it doesn't, try again with no clear reason. That approach occasionally produces a lucky result, but it rarely produces a repeatable one. Prompt craft treats prompt writing as a skill you can practice, measure, and improve. This article lays out an eight-step workflow for building prompts deliberately, shows a worked example, and explains how to use each tool's own controls. By the end, you should be able to get closer to the image you intended—not just a random image you happen to like.

1. Why a method beats inspiration

Inspiration is not a strategy. When you rely on a flash of insight, you cannot tell whether a good result came from your wording, the model's random seed, or pure chance. A method changes that. It gives you a fixed sequence of decisions—subject, medium, composition, light, style—so that when something works, you know which choice caused it. The difference is between a lucky first result and a reproducible workflow. With a workflow, you can return next week, change one element, and still get a usable image. You can also hand the prompt to a colleague and expect similar output. That reliability is what makes prompt writing a craft rather than a gamble.

2. The eight-step workflow

1. Write the subject in one plain sentence. Describe what is in the image without any style words yet. For example: "A barista pouring milk into a cup." This sentence is your anchor. If the model cannot render the subject, no amount of style language will save it.
2. Choose the medium or render style. Decide whether you want a photograph, an oil painting, a pencil sketch, or a 3D render. Naming the medium tells the model which visual rules to follow. A photograph implies realistic lighting and depth of field; a sketch implies visible lines and shading.
3. Set composition and framing. Specify how the subject sits in the frame: close-up, wide shot, from above, centered, or off to one side. Composition controls where the viewer's eye goes. Without it, the model often defaults to a centered, medium-distance view that may not suit your subject.
4. Describe the light. Light is the single most powerful mood tool. Say whether it is soft morning light, harsh noon sun, warm candlelight, or cool fluorescent overheads. Also note the direction: from the left, backlit, or from above. Consistent light direction makes an image feel intentional.
5. Add one style anchor, not five. A style anchor is a reference to a visual tradition or a named aesthetic—for example, "1970s film photography" or "Japanese woodblock print." One anchor gives the model a clear direction. Five anchors compete and often cancel each other out, producing a muddy result.
6. Generate and read the result like an editor. Look at the image and ask specific questions: Is the subject clear? Is the light consistent? Are there extra fingers or unreadable text? Write down what is wrong, not just whether you like it. This step turns a vague reaction into actionable notes.
7. Change exactly one variable and regenerate. Pick one item from your editor's notes—say, the light direction—and change only that in the prompt. Keep everything else the same. This isolates the effect of your change. If you change three things at once, you learn nothing about which one mattered.
8. Record the winning prompt and settings in a prompt log. A prompt log is a simple file or notebook where you save the exact prompt, the model name, the aspect ratio, the seed if available, and any other settings. Over time, this log becomes your personal reference. You will start to see patterns in what works for you.

3. One worked example

Suppose you want an image of a coffee shop. Here is a vague starting prompt: "A coffee shop scene." This might produce a generic interior with no clear focus. The model has to guess the subject, medium, composition, light, and style.

Version 2: "A barista pouring milk into a cup, photograph, close-up, soft morning light from the left." The change is the addition of a medium (photograph), a composition (close-up), and light direction. Typically, this version gives a more focused image with a clear subject and a consistent light source. The barista's hands and the cup become the center of attention.

Version 3: "A barista pouring milk into a cup, photograph, close-up, soft morning light from the left, 1970s film photography style." The change is one style anchor. This often adds a warm color cast and a slight grain, giving the image a specific mood. The subject and composition remain the same. As a general observation, each step tends to improve clarity and intentionality, but no outcome is guaranteed—models vary, and some generations will still miss.

4. Using each tool's own controls

Different image tools expose different controls. The Midjourney documentation describes parameters you can add to a prompt, such as aspect ratio (for example, --ar 16:9) and stylization (--stylize). These parameters let you control the shape of the canvas and how strongly the model applies its default aesthetic. The Google Gemini Apps Help Center explains how to revise an image through conversation—you can ask for changes in plain language, and the model updates the image iteratively. OpenAI provides guidance on its image tools, including how to describe scenes and how to use inpainting or outpainting to edit parts of an image. Each tool documents its own set of controls; reading those pages is the fastest way to learn what is possible. Do not assume that a parameter from one tool works in another.

5. Negative prompts, seeds, and settings on open models

Open models, such as those in the Stability AI ecosystem and community model cards on Hugging Face, often expose settings that closed tools hide. A negative prompt tells the model what to avoid—for example, "blurry, extra fingers, text." A seed is a number that initializes the random generation; using the same seed with the same prompt and settings tends to produce the same image, which helps you compare changes. Other common settings include step count (how many refinement passes the model makes) and guidance scale (how strictly the model follows your prompt). These are illustrative examples; exact names and ranges vary by model. Always check the model card for the specific controls it supports.

6. Quality checklist before publishing an AI image

Before you publish, run through this checklist. Anatomy and hands: Check for extra fingers, merged limbs, or distorted faces. Readable text: If the image contains text, verify it is spelled correctly and legible; AI models often produce garbled letters. Consistent light direction: Make sure shadows and highlights agree on where the light comes from. Brand or trademark elements: Look for logos or recognizable brand marks that you did not intend to include; remove or replace them. Disclosure of AI origin: Decide whether you need to state that the image was AI-generated. For a deeper discussion, see the copyright and provenance guide.

7. FAQ

Do I need to use all eight steps every time?

No. The workflow is a scaffold, not a rule. For a quick draft, you might skip the style anchor. But when you need a precise result, following all eight steps reduces guesswork.

What if the model ignores my negative prompt?

Negative prompts are suggestions, not commands. Some models weight them lightly. If a negative prompt is not working, try rephrasing it or adjusting the guidance scale if your tool exposes one.

How many style anchors should I use?

One. Adding more than one often creates conflicting visual cues. If you want to blend styles, do it in separate generations and combine them later in an editor.

Can I reuse a seed from another model?

No. Seeds are model-specific. A seed that works in one model will not produce the same image in another, even with an identical prompt.

Where can I learn more about prompt writing?

Start with the official documentation for your tool, then practice with the eight-step workflow. Our main guide covers the fundamentals of prompt structure.

Prompt craft is a skill you build over time. The eight-step workflow gives you a consistent way to improve, and the prompt log turns each attempt into a lesson. For more on structuring prompts, see the main guide, and for ethical and legal considerations, read the copyright and provenance guide.