A@creator
AI Prompt Templates
A prompt template is a prompt written to be reused. Instead of one sentence that produced one lucky image, a template names every variable it depends on — subject, product, palette, camera, layout, aspect ratio — and leaves those parts in [brackets] or in a JSON field so you can change them without rewriting the rest. This collection gathers the templated prompts in the gallery, including a large batch reverse-engineered from public GPT Image 2 results, where the structure of the prompt is the thing worth copying.
Templates matter because image models reward specificity but punish inconsistency. When you write freehand, you unconsciously change three things at once and cannot tell which one broke the render. A template freezes the parts that worked — lighting rig, lens, material language, typography rules — and exposes only the parts you actually want to vary. That is what turns a one-off image into a repeatable production recipe you can run across a whole catalog, a whole cast of characters, or a whole set of slides.
Every entry here shows the full prompt with its slots intact, the model label it was produced with, and a link back to the public source. Prompts that contain [bracket] variables open in the template workspace, where each slot becomes an editable field and the finished prompt is assembled for you to copy. Start with a template that is close to your brief, replace only the bracketed values, and change one structural line at a time so you always know which edit moved the result.
![Express [{argument name="subject" default="a powerful AI builder"}] in a… — GPT Image 2 prompt by @opc_8838](https://pub-fbc802e4b24043c4aab305c6666753cd.r2.dev/tweets/agi2_30/variants/v1/gpt-image-2-prompt-templates-prompt-1200.avif)
What separates a template from an ordinary prompt
The templates in this collection tend to share the same anatomy. Read one from top to bottom and you can usually see the four layers.
- Fixed scaffold: the sentence order, the sections, and the rendering vocabulary that must not change.
- Named slots: [subject], [product name], [brand color], [headline text] — the values you are meant to replace.
- Constraints: aspect ratio, resolution, what must stay legible, and what the model should not add.
- Negative guidance: the artifacts, watermarks, and extra limbs or extra text the prompt explicitly rules out.
JSON prompts and structured templates
Some of the strongest templates are written as JSON objects rather than prose: keys for subject, style, lighting, camera, composition and text, each with a short value. Models like GPT Image 2 and Nano Banana Pro follow that structure closely, and it makes the prompt trivially machine-editable — you swap a value instead of hunting through a paragraph. Prose templates remain better when the mood needs an unbroken sentence; use JSON when the brief is a checklist of attributes.
Frequently asked questions
- How do I use a prompt template?
- Copy it, then replace everything in [brackets] with your own values and leave the rest alone. In the gallery, prompts with bracket slots open a template workspace where each slot is an input field, so you can fill them in and copy the finished prompt without editing text by hand.
- Do prompt templates work across different image models?
- The structure transfers; the exact wording does not. Composition, lighting and subject lines survive a move from GPT Image 2 to Nano Banana or Midjourney, while model-specific parameters and text-rendering instructions usually need adjusting. Each entry shows the model it was captured with, so start there and adapt.
- What makes a template reusable rather than one-off?
- Variables that are named instead of assumed. If the prompt says "a red sports car" it is an example; if it says "[vehicle] in [brand color]" it is a template. The other half is constraint language — aspect ratio, legibility rules and negative guidance — which is what keeps results consistent between runs.
- Can I edit the fixed part of a template?
- Yes, but change one line at a time. The scaffold is the part that was tuned until the render was reliable, so an unmeasured rewrite is how templates stop working. Vary a single structural instruction, regenerate, and keep the version that holds up.