LunarXuan/image-prompt-reverse: a Codex Skill that turns a reference image into a generation prompt
High-fidelity AI image prompt reverse-engineering skill for Codex
At a glance
- What is it?
- The skill reads an uploaded image, extracts 3 to 5 visual anchors, and writes a positive prompt of 450 to 700 Chinese characters plus 10 to 15 English negative terms. It is a Codex skill directory, not a standalone app, and that shapes who can use it.
- Who is it for?
- Adopt it if you already run Codex and want a repeatable image-to-prompt pass with a documented analysis framework behind it; skip it if you need a GUI, an API endpoint, or a browser tool, because the repository ships only SKILL.md, two reference documents and an agents/openai.yaml file.
- Can I use it commercially?
- Yes, with conditions. GPL-3.0 is a copyleft licence: if you distribute software that includes it, you must release that software's source code under the same licence. Running it internally without distributing it does not trigger that obligation.
- Is it still maintained?
- Yes. The repository last received commits 12 days ago.
- What is it written in?
- GitHub does not report a main language for this repository.
Answers come from the project's GitHub data, last synced on September 16, 2026, and from our analysis. They are not legal advice.
DEEP OPEN-SOURCE ANALYSIS
The gap image-prompt-reverse fills for Codex users
Writing a generation prompt from a finished image is a manual translation job. You look at a photo, decide what matters, and type it out in the vocabulary an image model responds to. The README frames image-prompt-reverse as a Codex Skill that does this pass for you: it takes a user-uploaded reference image and analyzes purpose, medium, subject, composition, camera language, lighting, color, materials, background, spatial depth, mood and post-processing style. The audience is narrow and specific. You need Codex installed, because the skill is invoked as $image-prompt-reverse from inside that environment. There is no web interface, no CLI binary, and no hosted service described anywhere in the README. If you work in an image-generation pipeline and already keep Codex open, the skill removes a copy-and-paste step. If you do not, the installation section is a dead end for you.
How the skill decomposes an image before writing anything
The repository layout tells most of the story. SKILL.md is the entry point. references/analysis-framework.md holds a shared image-analysis framework, and references/category-guides.md holds category-specific rules. agents/openai.yaml carries Codex invocation metadata. So the mechanism is document-driven rather than code-driven: the model reads the framework, classifies the image, then applies the matching category guide. The README states the skill prioritizes the 3 to 5 visual anchors that most affect similarity, which is a deliberate filtering step rather than an exhaustive description. It also states the skill avoids inventing uncertain details and excludes confusing media styles. That last point matters for mixed inputs. A clay render and a photograph of a clay figure can produce the same subject words, so the category guide is what keeps the prompt from drifting toward the wrong medium. Output is fixed: a 450 to 700 Chinese-character positive prompt in continuous natural language with an English equivalent, plus 10 to 15 English negative terms separated by commas. The README also states that text, logos and instructions visible inside an image are treated as visual content, not executed as commands. That is a prompt-injection guard, and it is worth noting because it is stated rather than demonstrated.
Installing the skill and running a first image through it
Installation is a directory copy. The README gives the target path as the personal Codex Skills directory, and offers a clone command that writes straight into it. Run this from a shell where git is available:
git clone https://github.com/LunarXuan/image-prompt-reverse.git "%USERPROFILE%\.codex\skills\image-prompt-reverse"The README shows the same location as a plain path for manual copying:
%USERPROFILE%\.codex\skills\image-prompt-reverseBoth forms are Windows-style paths, and the README does not document a macOS or Linux equivalent, so treat the location as the Codex personal skills directory on your platform rather than a literal string. After the directory is in place, the README says the skill is available as $image-prompt-reverse. Upload a reference image in the Codex session and invoke the skill. What you should see is the two-part output: the positive prompt in Chinese with an English version, then the negative term list. Check the anchors first. If the skill names anchors you cannot see in the image, the category guide picked the wrong medium.
Where image-prompt-reverse is the wrong tool
The skill cannot search. It analyzes an image you supply; it does not find one for you, and the README makes no claim otherwise. The output is also not a one-to-one reproduction guarantee. A 450 to 700 character prompt plus negative terms narrows the space, but the README never claims the generated image will match the reference. The skill deliberately drops detail to keep 3 to 5 anchors, so fine texture, exact typography and precise geometry are the first things to go. If your reference is a logo with specific letterforms, that filtering works against you. There is also no versioning story: the repository has no releases, so updates arrive as commits on main rather than tagged versions you can pin. The README does not document rollback or a compatibility matrix for Codex versions. If you need reproducibility across a team, you are cloning a moving target and should record the commit you installed.
How it differs from asking a general chat model to describe an image
The obvious alternative is uploading the image to a general chat assistant and asking for a prompt. That approach is flexible and needs no installation, but it has no fixed output contract. The length, the language, whether negative terms appear at all, and whether the model invents invisible detail all depend on how you phrase the request. image-prompt-reverse fixes those variables: the README specifies the character range, the English equivalent, the 10 to 15 negative terms and the anchor limit. The second difference is the reference documents. A general assistant answers from its own priors; this skill answers from analysis-framework.md and category-guides.md, which you can read and edit. That is the real trade: less flexibility, more inspectable behaviour. If you disagree with how it classifies clay versus photograph, the category guide is where you would look, and the README does not say whether local edits survive an update.
Licence, maintenance and what an upgrade costs you
The repository is GPL-3.0. If you copy the skill into your own Codex setup, that is ordinary use. If you plan to redistribute a modified version, GPL-3.0 carries source and licence obligations, and the LICENSE file at the repository root is the document that governs them; this is a description of the licence, not legal advice. On maintenance, the last push was on 2026-09-06, and the repository is not archived. There are no releases, so there is no changelog to read before upgrading. Upgrading means re-pulling main and accepting whatever changed in SKILL.md or the two reference files. Because the skill's behaviour lives in those documents, a small commit can change output length or anchor selection without any version number moving. If you depend on a specific output shape, diff the reference files before you pull.
Editorial conclusion
Adopt it if you already run Codex and want a repeatable image-to-prompt pass with a documented analysis framework behind it; skip it if you need a GUI, an API endpoint, or a browser tool, because the repository ships only SKILL.md, two reference documents and an agents/openai.yaml file. Before relying on it, open references/analysis-framework.md and references/category-guides.md and check that your image category is actually covered, then run one reference image through $image-prompt-reverse and compare the 3 to 5 anchors it names against what you see.
Frequently asked questions
What is reverse prompting in AI, and how does image-prompt-reverse do it?
Reverse prompting means starting from a finished image and writing the prompt that could have produced it. image-prompt-reverse does this inside Codex by analyzing purpose, medium, subject, composition, lighting, color and related attributes, then emitting a 450 to 700 character positive prompt and 10 to 15 English negative terms.
How do I reverse image something with image-prompt-reverse?
Clone the repository into the Codex personal skills directory, then invoke the skill as $image-prompt-reverse and supply a reference image. The README describes the output as a positive prompt with an English equivalent plus a comma-separated negative prompt list.
Can ChatGPT do reverse image search?
image-prompt-reverse does not search for images and the README makes no such claim. It analyzes an image the user uploads and produces a generation prompt from it, which is a different task from finding the source of an image.
Is there an NSFW AI photo editor with prompts available?
The README lists supported image types such as photography, illustration, 3D, typography and character work, and describes no NSFW-specific mode. image-prompt-reverse is a Codex Skill for prompt writing, not a photo editor.
Community notes