OpenAI's Official 14-Point Prompt Guide for Consistent AI Images

You want to create images with the same atmosphere, but a different picture comes out every time.

When you line up three thumbnails, they look like they were made by three different people. You want to output the same person as before, but their face changes. Even if you type "same as last time," it doesn't seem to work at all.

This is the wall that most people who start using AI image generation for work hit first.

However, this isn't happening because your prompts are bad. The cause is much simpler: you're writing them from scratch every time.

Moreover, OpenAI has officially provided an answer to this problem. A template that breaks down image prompts into 14 items has been released. It includes 11 types for different uses, 8 types for editing, and completed examples.

Here are the four things I will share:

- The full text of the 14 official template items and what to write in each.
- A list of 11 types for specifying use cases in one word and 8 types for editing.
- The official method for continuing to output the "same face" and "same style."
- A way to avoid pasting that template every time (including a record of how I actually ran it).

The latter half is based on actual measurements. I passed one reference image, created a template, and went through the process of creating a 10-page document. I will write exactly how long it took, the fact that I redid it 7 times, and how I found out that you shouldn't use ImageGen directly for document creation.

First, here is the actual object. This is the official framework.

Official Original Text (from OpenAI imagegen skill's Shared Prompt Scaffolding)

``text
Use case: <taxonomy slug>
Asset type: <where the asset will be used>
Primary request: <user's main prompt>
Input images: <Image 1: role; Image 2: role> (optional)
Scene/backdrop: <environment>
Subject: <main subject>
Style/medium: <photo/illustration/3D/etc>
Composition/framing: <wide/close/top-down; placement>
Lighting/mood: <lighting + mood>
Color palette: <palette notes>
Materials/textures: <surface details>
Text (verbatim): "<exact text>"
Constraints: <must keep/must avoid>
Avoid: <negative constraints>
`

Source: [https://github.com/openai/skills/blob/main/skills/.system/imagegen/SKILL.md](https://github.com/openai/skills/blob/main/skills/.system/imagegen/SKILL.md)

Just looking at this, you might think "it's in English" and stop, so from here on, I will translate each item into Japanese and break down how to fill them out. Finally, I will provide a Japanese version of the template that you can copy and use as is.

## **The reason it varies every time is not because your prompts are bad**

The cause lies in the structure of the instructions, not the content.

You type "create a stylish blog header image" in free text. The next day, you type "simple and cool header image." In a human's mind, these are intended to be the same order, but these two are different instructions.

The former only has "stylish" as a clue. The latter has "simple" and "cool." The parts you don't specify are filled in by the AI each time. Because the way they are filled in is different every time, the pictures are different every time.

⚡ It's not the AI that's wavering; it's our instructions.

This is where the 14 items come in. If the items are fixed, you can see at a glance where you forgot to fill them in. If you open the previous file and only rewrite the lines you want to change, everything else will be the same as last time.

Stop "writing" prompts every time and change to "filling them in." That is the essence of this discussion.

**The official policy is "do not expand arbitrarily"**

There is another design philosophy that is helpful to know.

The official skill definition states:

`text
If the prompt is already specific and detailed, preserve that specificity and only normalize/structure it. If the prompt is generic, you may add tasteful augmentation when it will materially improve the result.
`

Source: [https://github.com/openai/skills/blob/main/skills/.system/imagegen/SKILL.md](https://github.com/openai/skills/blob/main/skills/.system/imagegen/SKILL.md)

If the instruction is specific, keep that specificity and just organize the structure. Only when the instruction is vague is it allowed to add elements to the extent that the result improves. That's how it's decided.

In other words, the more detailed you write, the less discretion the AI has. Less discretion means less variance.

Conversely, for vague instructions, the AI is officially permitted to "be helpful and add things." The parts added out of goodwill are different every time. This is the source of the variance.

📌 Therefore, the 14 items are tools for giving information to the AI, and at the same time, tools for taking away the AI's freedom. If you view it that way, you won't use it incorrectly.

## **Before that: You probably already have this tool**

No installation is required.

imagegen is a "system skill" that has been in Codex from the beginning. It's not something users add later; it's ready to use the moment it's started. No additional contracts or API keys are required.

In the official description, the default behavior is a mode that uses the built-in image_gen tool, which looks like this:

Built-in tool (default): Does not require OPENAI_API_KEY.

An API key is only needed for the CLI fallback (scripts/image_gen.py) that runs only when explicitly requested. For normal use, it doesn't matter.

⚠️ This is where 90% of people stumble: The calling symbol is different depending on the location.

If you don't know this, you'll get stuck on the first line. The official documentation describes it as follows:

`text
In Codex CLI or the IDE extension, run /skills or type $ to mention a skill.
`

`text
In ChatGPT Work, use @skill-creator format.
`

Source: [https://learn.chatgpt.com/docs/build-skills](https://learn.chatgpt.com/docs/build-skills)

In Codex CLI and IDE extensions, it's $. In ChatGPT Work, it's @.

Even though it's the same function, the calling symbol is different. When the procedure you found online doesn't work, it's usually just that the screen you're on is different from the screen in the article. The grammar isn't wrong.

The official documentation also lists where it can be used:

`text
Standalone skills are available in the ChatGPT desktop app, Codex CLI, and IDE extension. Skills bundled in plugins are also available in Chat and Work across ChatGPT on the web, desktop, and mobile.
`

Source: [https://learn.chatgpt.com/docs/build-skills](https://learn.chatgpt.com/docs/build-skills)

Desktop app, Codex CLI, and IDE extension. These three are the arena for standalone skills.

🎯 Even people who can't use the skill can use the template starting today.

This is an important point, so I'll state it clearly. For people who don't use Codex or the desktop app, imagegen itself is irrelevant. If you just open ChatGPT in a browser, nothing will come up if you type $.

However, the 14-item template is just a text format. You can paste it directly into the image generation input field of browser-based ChatGPT, Gemini, or Claude. The skill is just a "mechanism that writes in this format for you," and the format itself can be used by anyone.

Therefore, the value of this article does not change depending on whether you can use the skill. Those who can use it should take it as automation, and those who can't should take it as a copy-paste template.

## **Prompt Template: All 14 Items**

From here is the main body. Let's look at what each field is for, in order. There are 14 items, but you don't need to fill them all. It's just a relationship where the AI complements the parts you didn't fill in. It's enough to fill in only the items you're having trouble with variance.

1. Use case
This is the field to specify the genre of the image using a fixed word. Instead of free words, you choose from words prepared by the official (taxonomy slug). There are 11 types in total, such as product-mockup, ui-mockup, and logo-brand. A list is summarized in a later chapter. Deciding this first automatically determines how to fill in the subsequent items. If it's a product photo, you need lighting. If it's a diagram, you need text. Once the use case is decided, the necessary fields are also decided.

2. Asset type
Write where the image will ultimately be placed. "Blog header image," "Landing page hero background," "Game UI icon." If the use case (No. 1) is a major category, this is the actual installation location. Even for the same photo-style image, the required margins and aspect ratio change depending on whether it's placed in a header or an icon. This is the field to convey that.

3. Primary request
This is the field to put the "make that" that you usually type in free text. In the official explanation, it's <user's main prompt>. In other words, these 14 items are not a mechanism for discarding free text. It's a structure where you place the free text in the No. 3 field and line up conditions around it. If this is empty, nothing starts. Conversely, if you only fill this in, it's the same state as before.

4. Input images
When passing images you have on hand, this is the field to write what role each one plays. The official format is <Image 1: role; Image 2: role>. You declare the number and role as a set, like "Image 1 is a person photo, Image 2 is a clothing reference." Since it's an optional item, you can leave it empty when creating from scratch. However, when you want to keep outputting the same person, this field becomes the protagonist. More on this in the editing chapter.

5. Scene/backdrop
The environment where the picture is placed. "On a wood-grain table," "Inside an industrial hangar," "Light gray with a slight texture." It can be a real place or an abstract background. If you don't write the background, it will change every time. The reason the atmosphere of thumbnails doesn't match is usually because this field is empty.

6. Subject
What to place in the center of the screen. You can write not only the object itself but also how it's placed, like "A single bottle placed in the center, with a slight reflection." The point is to write the background (No. 5) and the subject (No. 6) separately; if you combine them into one sentence, the AI won't be able to judge which to prioritize.

7. Style/medium
Is it a photo, an illustration, or 3D? In the official examples, it's written at a granularity like "matte illustration," "cinematic concept art," or "vector-style diagram." "Stylish" and "cool" are not words to put here. Think of it as a field to write the name of the technique.

8. Composition/framing
Is it a wide shot or a close-up, from directly above, and where to place the subject? The official example has a way of writing like "wide composition with usable negative space for page copy." This means a horizontal composition with margins left on the left and right for placing text. This field is effective for images where you will add text later, like thumbnails or headers. If you don't reserve the margins first, the subject will come to the full screen and the text won't fit.

9. Lighting/mood
Is it soft morning light, studio lighting, or light rays in the fog? The official examples use expressions like "soft morning light," "softbox lighting, clean highlights, controlled shadows," and "volumetric light rays cutting through fog." If you want to stabilize the photo-likeness, the fastest way is to fix the same wording here every time.

10. Color palette
The policy for the colors to use. "Brand colors," "subdued neutral colors," "deep green and off-white." It can be specific color names or a direction. If you want to show brand consistency, write your colors here and fix them. This alone will align the impression when they are lined up.

11. Materials/textures
The texture of the surface. Write information related to how light bounces, like "matte plastic, clearly printed label." This field is effective for product photos and can be omitted for landscapes or illustrations without issue.

12. Text (verbatim)
A field to write the text to be included in the image exactly as it is. The point is that the official format is Text (verbatim): "<exact text>", and verbatim means "exactly as in the original." You pass the exact string you want to output, enclosed in quotation marks, not a summarized content. In the official example for a diagram, it's written like this:

`text
Text (verbatim): "Bean Hopper", "Grinder", "Brew Group", "Boiler", "Water Tank", "Drip Tray"
`

Source: [https://github.com/openai/codex/blob/main/codex-rs/skills/src/assets/samples/imagegen/references/sample-prompts.md](https://github.com/openai/codex/blob/main/codex-rs/skills/src/assets/samples/imagegen/references/sample-prompts.md)

We often hear that text in AI images gets garbled, but in many cases, the exact text to be included wasn't passed correctly in the first place. Just filling this in can change the result.

13. Constraints
A field to list "keep this" and "don't do this." The most common set in official examples is "no text; no logos; no watermark." For images intended to have text added later, including these three every time reduces accidents. It stops the AI from helpfully adding English catchphrases on its own.

14. Avoid
While No. 13 is for general conditions, this field is narrowed down to "things you don't want to output." Gradients, 3D, being placed inside a mockup. In the official example for a logo, it's written as "no gradients; no mockups; no 3D; no watermark." Since the roles overlap with No. 13, often only one is used in actual official templates. Rather than strictly distinguishing them, it's enough to understand that "there are two places to write things you want it to stop doing."

## **Specify the use case in one word. 11 types prepared officially**

This is a list of words to put in No. 1 Use case. In the official sample collection, 11 types are listed for generation.

- photorealistic-natural (realistic photos, natural scenes)
- product-mockup (product photos, packaging)
- ui-mockup (screen design mockups)
- infographic-diagram (diagrams, infographics)
- scientific-educational (diagrams for science and education)
- logo-brand (logos, brand marks)
- illustration-story (illustrations for stories)
- stylized-concept (stylized concept art)
- ads-marketing (advertising and marketing materials)
- productivity-visual (visuals for documents and business)
- historical-scene (historical scenes)

In addition to these, templates by installation location are also prepared. For websites (hero images, section illustrations, blog headers), for games (background concepts, characters, UI icons, textures), wireframes (PC, pricing pages, smartphone onboarding), logos (abstract marks, monograms, wordmarks). For blog header images, the official example is provided as is.

Official Original Text (Website assets example: blog header image)

`text
Use case: photorealistic-natural
Asset type: blog header image
Primary request: overhead desk scene with notebook, pen, and coffee cup
Scene/backdrop: warm wooden tabletop
Style/medium: photorealistic photo
Composition/framing: wide crop with clean room for page copy
Lighting/mood: soft morning light
Constraints: no text; no logos; no watermark
`

Note that only 8 out of 14 items are filled in. The official itself doesn't fill them all in. It serves as an example that you only need to fill in the necessary fields.

## **Continuing to output the same face and style. 8 types for editing**

This is the direct answer to "it varies every time." In the official sample collection, there are 8 types of use cases for editing existing images.

- text-localization (replacing only text, such as translation)
- identity-preserve (changing while maintaining the identity of the person)
- precise-object-edit (replacing only specific objects)
- lighting-weather (changing only light and weather)
- style-transfer (transferring the style of a reference image)
- compositing (compositing subjects from another image)
- character consistency workflow (outputting the same character in a different scene)
- sketch-to-render (turning a sketch into a final picture)

The one that works for the face-changing problem is identity-preserve. The official example is as follows:

Official Original Text (Edit: identity-preserve)

`text
Use case: identity-preserve
Input images: Image 1: person photo; Image 2..N: clothing references
Primary request: replace only the clothing with the provided garments
Constraints: preserve face, body shape, pose, hair, expression, and identity; match lighting and shadows; keep the background unchanged; no accessories or text
`

The key part is the Constraints line. Preserve face, body shape, pose, hair, expression, and identity. Match lighting and shadows. Keep the background unchanged. In other words, instead of saying "make it the same person," it lists the things you want to be the same by name. This way of writing is the official answer.

🔥 If you're outputting many characters, use this format

For uses like series thumbnails or outputting the same person many times, there is a more detailed format.

Official Original Text (Edit: character consistency workflow)

`text
Use case: identity-preserve
Input images: Image 1: previous character anchor illustration
Primary request: continue the story with the same character in a new scene and action
Scene/backdrop: snowy forest after a winter storm
Subject: same young forest hero gently helping a frightened squirrel out of a fallen tree
Style/medium: same children's book watercolor illustration style as Image 1
Constraints: do not redesign the character; preserve facial features, proportions, outfit, color palette, and personality; no text; no watermark
`

The term "character anchor" used here is the key. You specify one image that serves as the standard as an "anchor," and from then on, you only instruct the differences from there. The point that the Style/medium field is "same children's book watercolor illustration style as Image 1" is also important; it doesn't explain the style in words. It specifies it by reference as "same as Image 1." Trying to reproduce a style with words always leads to deviation. Pointing by reference is more accurate.

If you only want to align the style, use style-transfer

If you want to align the overall tone rather than the person, use this.

Official Original Text (Edit: style-transfer)

`text
Use case: style-transfer
Input images: Image 1: style reference
Primary request: apply Image 1's visual style to a man riding a motorcycle on a plain white backdrop
Constraints: preserve palette, texture, and brushwork; no extra elements
`

Preserve palette, texture, and brushwork. Don't add extra elements. If you have one image you like on hand, place it in Image 1 and then swap the content. The brand's tone and manner will be most stable in this format.

## **Japanese version of the template that can be used as is**

The following two are not the prompts themselves distributed by the official. They are reconstructed into a form that can be filled in in Japanese based on the official item structure cited above. The item names retain the official English labels.

★ Application (Created based on the official item structure. Not a prompt distributed by the official)

`text
Use case:
Asset type:
Primary request:
Input images:
Scene/backdrop:
Subject:
Style/medium:
Composition/framing:
Lighting/mood:
Color palette:
Materials/textures:
Text (verbatim):
Constraints: no text; no logos; no watermark
Avoid:
`

Create just one of these in a text file and keep it. From next time, open this and rewrite only the lines you want to change. If you fill it in for a thumbnail, it will look like this:

★ Application (Based on the official Website assets / blog header example, reconstructed in Japanese)

`text
Use case: photorealistic-natural
Asset type: Blog article header image (horizontal)
Primary request: Composition looking down at a desk from diagonally above. Notebook, pen, and mug
Scene/backdrop: Warm wood-grain desk
Subject: Notebook slightly left of center, mug in the right foreground
Style/medium: Realistic photo
Composition/framing: Horizontal. Leave margins on the right to place the title
Lighting/mood: Soft natural morning light
Color palette: Calm color scheme centered on wood brown and off-white
Constraints: no text; no logos; no watermark
`

From the second time, rewrite only the two lines: Primary request and Subject. Keep everything else as it was last time. This will align the "same atmosphere as before." The trick is to decide which lines to fix and which to change at the beginning. I think the allocation of fixing the four lines of background, style, light, and color palette, and moving only the subject and composition, is the easiest to handle.

## **If pasting the template every time is a hassle, you can create the whole template**

From here on are the records of what I actually ran. Honestly, it's a pain to copy and paste the 14 items every time. A mechanism to solve that was prepared on the Codex side. When you add Image Gen to the input field, a container called the Template Gallery appears. You can place your own templates here. When creating it, I asked the Template Creator like this:

★ Application (Instructions I actually sent. Not a prompt distributed by the official)

`text
Please use the Template Creator to create a new image template for ImageGen. First, explain how the template works and how to use it. Next, request an upload of a PNG reference image and ask me when and how to use the template if necessary.
`

The response came back in 27 seconds. It was good that it asked two things first instead of starting the work suddenly. The first was "Do you want the template to lean towards the design of the entire slide, or do you want to use it for generating illustrations and diagrams?" The second was "Is the main use closer to a research report, seminar materials, sales materials, or an SNS carousel?" These two questions are the branch points for the design. Even with the same "wanting to align the design," the resulting template changes depending on whether you're creating a full visual or a picture intended to have text placed in the margins. It makes sense to have this decided first.

⚠️ The true identity was not "template registration" but "skill saving"

I was also misunderstood until halfway through, so I'll state it clearly. It's not that there's a template registration mechanism like Canva inside ImageGen. According to Codex, it was actually a mechanism where rules for reference images, color palettes, expressions, and layouts are saved as a "skill" in Codex and called from ImageGen. A skill is a reusable workflow that combines instructions, reference materials, scripts, etc. The official explanation is here.

Source: [https://developers.openai.com/plugins/concepts/skills.md](https://developers.openai.com/plugins/concepts/skills.md)

Therefore, the act of creating a template is essentially growing one skill dedicated to yourself. If you switch to this understanding, the next part makes sense.

How to call the created template

When completed, the template is named and put in the gallery. From next time, you can call it in two ways:

- Add Image Gen to the input field and choose from the gallery
- Directly specify the ID starting with $artifact-template-

You can also call it by name. Writing "Create with the design of (template name)" worked. 📌 In other words, the 14 items are a "format written by hand," and the template is "that format saved as a skill." As for the order, it's safer to write by hand first to grasp the format and then turn it into a template. If you create a template suddenly, you won't be able to explain what you fixed yourself.

## **If you're making documents, you shouldn't use ImageGen directly**

This was the most important discovery this time. Now that the template was ready, I tried to have it create a 10-page document. As a result, it was able to create it, but when I asked Codex itself "what is the correct way to ask in the future," I got an answer different from what I expected. Using ImageGen directly is suitable for single image production. Specifically, when you want to make one illustration, when you want to make only a cover image, when you want to fix the color or composition of an existing image, or when you want to output one design proposal as a trial. Conversely, it's not suitable for the production of the entire document. The reason was also clear. It's because it's difficult to collectively manage research, 10-page structure design, unity between pages, and conversion to PowerPoint and PDF.

🎯 ImageGen is a tool specialized for one image. Leave the document to the skill. If you have it create a document without knowing this line, it will be a detour.

The correct way to ask is "instruct the whole with a skill and have it use ImageGen within that"

The division of roles was as follows:

- Design template (the one you made yourself): Reproduces color palette, decoration, and worldview
- Skill for slide generation: Responsible for research, structure, image generation, verification, and conversion to PPTX and PDF
- ImageGen: Actually generates the completed images for each page

If you specify the top two and send it, ImageGen is called internally. This is the way of writing that was actually guided:

★ Application (The form of specification guided by Codex. The theme part is hidden and generalized)

`text
Using $imagegen-slide-deck and $artifact-template-(your template name), please create a 10-page document on the theme of "XX." With the same design as this time, please save the PowerPoint and PDF with the official title as the file name.
`

If only the reproduction of the design is enough, the template side alone is sufficient.

★ Application (Same as above)

`text
Using $artifact-template-(your template name), please create a 10-page slide on the theme of "XX." Research the content and save it as PowerPoint and PDF.
`

⚡ The resulting PPTX cannot have its text edited

If you don't know this, an accident will surely happen. The PowerPoint that comes out in this way has each page as a completed image. In other words, even if you open the PowerPoint, you cannot directly rewrite the text. There is a workaround, and I was told that if I clearly state "in a PowerPoint with editable text" when requesting, it will be fine. For documents that you want to fine-tune yourself before distributing, please add this from the beginning. Attention is also needed for the file name. If left alone, it was saved with a generic name like deck.pptx. I make sure to write "save with the official title as the file name" every time.

Have the source included in the presenter notes

This was subtly effective, so I'll write it down. I was able to have the URLs of the researched sources put in the PowerPoint presenter notes. Since I can follow "where did I get this description from" later, I think it's an essential practice for documents distributed to others. When I asked for a separate fact check, it returned the answer by listing the URLs of the evidence and separating the parts that were not proven and the parts where interpretations differed by school or position. This was a part that was honestly helpful.

Time taken

Actual measurement. It went through 7 rounds in total. The first template design consultation was 27 seconds. From passing the reference image to creating the template body and the first version of 10 pages took 23 minutes and 9 seconds. The second time, when I rebuilt the structure, took 13 minutes and 13 seconds. Fixing only one page took 11 minutes and 34 seconds. Fact-checking and correction took 14 minutes and 18 seconds. Changing the wording once more took 6 minutes and 52 seconds. Saving and checking the template took 3 minutes and 11 seconds. Organizing the file names took 1 minute and 10 seconds. In total, about an hour and a half. However, this was not the time I was glued to the screen, but the time I spent waiting after sending. I was doing other work in the meantime.

## **Stumbling blocks found after redoing 7 times**

This might be what I want to convey the most. It wasn't completed in one go. I went through 7 rounds. The first version came out, I fundamentally rebuilt the structure, fixed only one page, fact-checked, changed the wording, saved, and organized the file names. The fact that I could "fix only one page" was big, and it didn't become a complete redo.

Stumbling block 1: The first version comes back as "just lined up"
The first 10 pages that came out were correct as information, but weak as a document. The items were just lined up, and it wasn't a flow that would convince the reader. So for the second time, I changed the way I ordered. Instead of specifying "create in this order," I sent this:

★ Application (The gist of the instructions I actually sent. Not a prompt distributed by the official)

`text
To avoid failing with the slides, I'll leave it to you to decide what kind of structure and flow to proceed with, so please create it in the best way possible.
``

Don't specify the structure, leave the design of the structure itself to the AI. The second time that came back, the format of each page was unified, an explanation of the mechanism was included in the middle, and a cross-reference quick-reference table was added in the latter half. It was a structure I wouldn't have thought of myself. It's not necessarily the case that the more detailed you instruct, the better it gets. The requirements for the content should be detailed, and the structure should be left to the AI. I think this allocation was good.

Stumbling block 2: Text gets on top of the picture and becomes unreadable
The most obvious failure was this. The annotation put at the bottom of the page overlapped with the background illustration and was completely unreadable. The fix was simple: I moved the annotation from on top of the illustration to the very bottom inside the white panel. This eliminated the interference with the background. This is the same story as specifying "leave margins to place text" in the 14 items' Composition/framing. The structure where the picture and text are competing for the same place doesn't change whether it's a single image or a 10-page document.

Stumbling block 3: The last page is weak
The final page of a document is usually the place to lead to an inquiry. In the first version, this was just one word and was too simple. What I fixed was dividing one sentence into three stages. Convey that there is such an opportunity now. Encourage them to let me know their situation or what they are having trouble with. Then show the method of contact. If it's just "please contact me," the reader has no reason to move. Reason, an atmosphere that's okay to talk in, and method. Just by placing these three separately, the appearance of the final page changed.

Stumbling block 4: It asserts things that shouldn't be asserted
This is a story that works regardless of the field, so I'll write it at the end. In the first version, there were descriptions that asserted even the range that shouldn't be asserted. Things that are traditionally said and things that can be objectively confirmed are different, but they were written in the same tone. So, I had it write separately what can be confirmed as facts and what is just interpreted as such. It avoids assertions and clearly states what kind of nature the document has. 🔥 This is a part that only humans can see. AI is good at organizing the appearance as a document, but it doesn't judge the line of how far it's okay to assert. The more it's a document to be distributed, the more you should reread this part yourself.

## Cross-reference: Find items from what you're having trouble with

Finally, I'll summarize it in a form that can be pulled from symptoms.

- The picture becomes completely different every time → Fix the four lines: Scene/backdrop, Style/medium, Lighting/mood, and Color palette
- A person's face changes every time → Set Use case to identity-preserve and place one standard image in Input images
- Want to output the same character in a different scene → Use the character consistency workflow format and specify Style/medium as "same as Image 1"
- Want to align only the style → In style-transfer, place one image you like in style reference
- No margins to place text → Write the specification for margins in Composition/framing
- English text you didn't ask for is included → Put "no text" in Constraints
- Text in the image is garbled → Pass the string you want to output in Text (verbatim), enclosed in quotation marks
- The texture of product photos is not stable → Fill in Materials/textures
- No response even if typed according to the procedure → Check the screen you're on. Codex CLI and IDE extensions are $, ChatGPT Work is @
- Pasting the template every time is a hassle → Turn it into a template with Template Creator and call it from the gallery
- Want to create a multi-page document → Don't use ImageGen directly; specify both the slide generation skill and the template
- Cannot rewrite text in PowerPoint → Clearly state "in a PowerPoint with editable text" when requesting
- File name becomes a generic name like deck.pptx → Write "save with the official title as the file name" every time
- Want to align the slide layout as well → Pass not only the reference PNG but also the original PPTX together
- Text in the document overlaps with the picture → Have the text placed inside the white panel instead of on top of the picture
- The resulting document is "just lined up" and weak → Don't specify the structure, leave the design of the structure itself to the AI again

## One thing to do today

If you try to do everything at once, you probably won't continue. First, decide on just one type of image you create most often. Whether it's a blog header, a thumbnail, or an illustration for a document. Just fill in the template for that one type in the blank template above and save it. That's it. You don't have to fill in all 14 items. If the four lines of background, style, light, and color palette are included, the picture you make next time will already lean towards the previous one. It's fine to proceed to templating or slides after that. If you reverse the order, you'll end up automating without knowing what you fixed yourself. First, fill in one by hand. Use it repeatedly and notice the lines that are missing. Then turn it into a template. Change from the job of writing prompts to the job of filling in prompts. When this changes, the content of the time spent on image creation will be replaced. Thinking time decreases, and choosing time increases. Try opening the one you made next week. The place where you felt a line was missing is the item necessary for your work.

References/Sources

- OpenAI Official imagegen skill definition (Shared Prompt Scaffolding / Default mode and necessity of API key / Policy for prompt completion)
https://github.com/openai/skills/blob/main/skills/.system/imagegen/SKILL.md

- OpenAI Official imagegen sample prompt collection (11 types for generation / 8 types for editing / Templates and examples by installation location)
https://github.com/openai/codex/blob/main/codex-rs/skills/src/assets/samples/imagegen/references/sample-prompts.md

- OpenAI Official documentation "How to call skills and where they are available" (Difference between $ and @ / System skills)
https://learn.chatgpt.com/docs/build-skills

- OpenAI Official documentation "Skills" (Mechanism where the true identity of a template is saved as a skill)
https://developers.openai.com/plugins/concepts/skills.md

- Template Gallery, Template Creator, the flow from research to document generation, time required, and stumbling blocks are records of the author's actual operations in August 2026. The screen configuration and names may change in the future.

14 Rules to Reduce Image Generation Variance Released by OpenAI (Prompt Included)

Turn one viral article into a full content workflow

Recent viral articles

RICOCHET ANTI-CHEAT™ UPDATE: More Experts. Stronger Defenses. Fairer Play.

Someone is making 300+ SOL by LPing on Meteora. You can just copy his positions.

This New App Is Giving Hyperliquid Some Serious...‘Fomo’

Update on the upcoming OpenClaw release

I Joined Midacy Co., Ltd.: My New Challenge in Men's Grooming

Hiroshi Sugimoto is Tamori in the Guise of Daisetz Suzuki: Reflections on the "Extinction Photographs" Exhibition