A prompt is a plain-language instruction, and the more specific it is, the closer the output gets to what you wanted, because the model only knows what you wrote.
- Be specific. Name the camera, lens, lighting and style; each named detail removes one thing that could go wrong.
- Follow a simple structure. Role or subject first, then context and environment, then constraints and format.
- Check for a negative field. Midjourney has one; Veo has one only on Google’s Gemini Enterprise Agent Platform; FLUX.2, Runway and Seedance do not.
- Save the seed. The seed changes every run by default, so reuse it to reproduce a result you like.
- Expect to iterate. Three to five rounds of refinement is normal for commercial-grade output.
Written for beginners and still accurate. When you are ready for what changed since: context engineering covers why prompting is now the smaller half of the job.
A prompt is an instruction. You write it in plain language, send it to an AI model, and the model generates an output based on what you wrote. That's the whole thing. The gap between a bad output and a great one almost always comes down to the quality of the instruction — not the model.
This guide covers everything you need to start writing prompts that work. No technical background required.
What a Prompt Actually Is
A prompt is everything the AI model receives before it starts generating. That includes your main instruction, any context you provide, and any constraints you set. Unless memory is switched on in the product you are using, it carries nothing over from earlier sessions. It has no understanding of what you intended if you didn't say it, and no way to ask follow-up questions — unless you build that into the prompt itself.Some products now remember by default: Anthropic, Claude's memory works everywhere, and you decide what's in it, 25 Aug 2026, read at source 16 Sep 2026: “Memory is on by default on Free, Pro and Max plans across web, desktop, and mobile.” An earlier version of this paragraph said the model has no memory between sessions.
This is the core insight that changes how most people approach prompting: the model doesn't know what you mean. It only knows what you wrote. Every detail you leave out is a detail the model will fill in on its own — and it will almost always fill it in wrong for your specific purpose.
The model only knows what you wrote. It cannot ask questions, so it fills in anything you leave out, usually wrongly for your purpose.
The Single Most Important Concept: Specificity
Vague prompts produce generic outputs. The more specific your instruction, the closer the output is to what you actually wanted. This is true for every type of AI model — image generators, video generators, and LLMs all behave this way.
SPECIFIC PROMPT: "RAW photo, 1woman, late 20s, ivory blazer, sitting at a marble cafe table, golden hour light from left window, natural skin texture, Canon EOS R5 85mm f/1.2, Vogue editorial style, 8K"
Same subject. Completely different output quality.
The specific prompt names the camera. Names the lighting. Names the publication reference. Names the lens. Every named element removes one degree of freedom from the model's interpretation — and reduces one variable that could go wrong.
Vague prompts give generic outputs. Every named element, such as camera, lens or lighting, removes one degree of freedom and one variable that could go wrong.
The 3-Part Structure Every Prompt Needs
Most professional prompts follow a simple structure, whether or not the person writing them knows it:
1. Role or Subject
Who or what is the subject of the output? For image prompts: describe the person, object, or scene. For LLMs: define the role the AI should take ("You are a senior brand copywriter for a luxury fashion brand"). Put it early. Order matters on at least some models, though vendors differ on what goes first: Google’s Veo formula opens with the camera, for example.Black Forest Labs, FLUX.2 prompting guide, read at source 17 Sep 2026: “Word order matters - FLUX.2 pays more attention to what comes first.” An earlier version of this paragraph said all AI models weight earlier tokens more heavily; no source says that for every model.
2. Context and Environment
Where is this happening? What's the setting, the mood, the time of day? For image prompts: describe the environment, lighting conditions, and atmosphere. For LLMs: describe the situation, the audience, and any relevant background the model needs to know.
3. Constraints and Format
What do you want the output to look like? For image prompts: camera, lens, color grade, resolution, style references. For LLMs: word count, format (bullets vs prose), tone, things to avoid. Constraints narrow the model's output space and make results more consistent across multiple runs.
Subject early, then context, then constraints. Order matters on some models, so check your vendor’s formula, and constraints like format, tone or lens make results more consistent across runs.
Negative Prompts
A negative prompt is a list of things you don't want in the output. Check whether your model has one before you write it — this is the detail most beginner guides get wrong. Midjourney has a negative field (--no); Veo has negativePrompt only on Google’s Gemini Enterprise Agent Platform, not in the Gemini API; and Kling has one only on its legacy API — where Kling itself recommends writing negatives inside the prompt. FLUX.2, Runway, Seedance and ChatGPT Images do not, and Runway states that using one “may result in the opposite happening.”Kling AI API, Text to Video (legacy), read in a browser 16 Sep 2026: “It is recommended to supplement negative prompt via negative sentences within positive prompts”. An earlier version of this paragraph listed Kling alongside Midjourney and Veo without qualification. Where the field exists, it is worth using. Where it does not, the same thinking still applies — you describe your way past the failure instead of naming it. Which model does what.
On a model that reads one, a negative prompt names the failure modes to avoid. On a model that does not, the same list is still worth writing — as a checklist of what your positive prompt has to rule out. Instead of “no plastic skin” you write “visible pores, uneven natural skin tone.” Same intent, phrased so the model can act on it.
Not every model reads a negative prompt. Where the field exists, use it. Where it does not, describe your way past the failure, such as asking for visible pores instead of banning plastic skin.
Why the Same Prompt Gives Different Results Each Time
AI models use randomness as part of the generation process. A setting called the seed controls this randomness — and by default, it changes every run. That's why two runs of the same prompt produce different outputs. If you find a result you like and want to reproduce it, save the seed number and use it again. Most image models display the seed in the output metadata.
The seed controls randomness and changes every run by default. Save the seed of a result you like so you can reproduce it.
The Iterative Workflow
Professional prompt engineers don't write one prompt and expect a perfect output. They write a base prompt, review the output, identify what's wrong, and refine. ChatGPT Images 2.5 makes this particularly straightforward — you can describe changes conversationally ("make her jacket darker, change the background to an interior") and the model adjusts while keeping everything else the same.
For any other model, iteration means updating the prompt text and regenerating. Three to five rounds of refinement is normal for commercial-grade output. First-run perfection is rare and shouldn't be expected.
First-run perfection is rare. Review, spot what is wrong and refine; ChatGPT Images 2.5 lets you do this conversationally, while other models need the prompt text updated.
Where to Go Next
This article covers the foundation. The rest of the site goes deeper into every area:
- What makes a prompt good — the elements every working prompt shares.
- The iterative workflow — how to refine an output instead of starting over.
- What the vendors actually say — each vendor’s own prompt formula, side by side.
- Model prompting cheat sheets — one sheet per model, including whether it takes a negative prompt.
- Prompting for LLMs — what differs between ChatGPT, Claude and Gemini.
More prompts, when something changes.
Prompts, model guides and workflow notes, sent when there is something new. Free.
SUBSCRIBE FREE ↗