# image-prompt-reverse is four files of instructions, not a model > This Codex skill reads an uploaded reference image along twelve dimensions, picks the three to five visual anchors that drive similarity, and writes a 450 to 700 character Chinese prompt plus a negative prompt. It ships no code, no tests and no POSIX install path, and the high-fidelity claim has nothing in the repository to measure it. **LunarXuan/image-prompt-reverse** — High-fidelity AI image prompt reverse-engineering skill for Codex - Repository: https://github.com/LunarXuan/image-prompt-reverse - Stars: 460 · Forks: 35 - Language: Unknown - License: GPL-3.0 - Published: 2026-09-17 · Updated: 2026-09-17 · Language: en - Canonical page: https://hysenlabs.com/projects/lunarxuan-image-prompt-reverse ## Four files, no package.json, no tests The repository root holds `LICENSE`, `README.md`, `SKILL.md`, `agents/` and `references/`. There is no package manifest, no dependency file, no build step and no test directory, and it publishes no GitHub releases, so there is no version to pin and no changelog to read. What you install is a set of instructions that a coding agent reads when you invoke it, which is why the only configuration surface described anywhere is a single YAML file and a name typed at the prompt. The last push was on 2026-09-06, and the repository lists no homepage, so there is no site to check the version against. The README is bilingual, with a Chinese section first and an English section after it, each carrying the same supported types, output contract and install steps, which is generous for readers of either language and doubles the maintenance surface for whoever edits it. ## The documented install path is a Windows environment variable Installation means putting the repository directory inside Codex's personal skills directory, and the only path given is a Windows one: ```text %USERPROFILE%\.codex\skills\image-prompt-reverse ``` Or clone it there directly: ```bash git clone https://github.com/LunarXuan/image-prompt-reverse.git "%USERPROFILE%\.codex\skills\image-prompt-reverse" ``` No POSIX equivalent appears anywhere, so on Linux or macOS you are left inferring where the equivalent directory lives, and the README does not say. After installation the skill is invoked explicitly with `$image-prompt-reverse`, which means it does nothing until you ask for it by name rather than reacting to any image you attach. ## Twelve dimensions, then three anchors worth keeping The analysis frame is explicit about what it reads from an image: purpose, medium, subject, composition, camera language, lighting, color, materials, background, spatial depth, mood and post-processing style. Supported types follow the same enumerative style, from photography (portraits, products, documentary, food, animals, nature, cityscapes) through illustration (flat, anime, hand-drawn, cyberpunk) and 3D (realistic, clay, designer toys, product rendering) to typography, logos, posters and IP-inspired or chibi characters. What makes the frame usable rather than a checklist is the prioritisation rule: the skill concentrates on the three to five visual anchors with the greatest effect on similarity, avoids inventing details it cannot see clearly, and drops media styles that could be confused with the target one. Category-specific rules live separately in `references/category-guides.md`, which is the file to read before trusting the output on an unfamiliar kind of image. ## 450 to 700 Chinese characters, and an English version of unstated length The output contract is two sections with hard numbers on one side. The positive prompt is a continuous natural-language passage of 450 to 700 Chinese characters, followed by an equivalent English version. The negative prompt is 10 to 15 English terms separated by commas. Note the asymmetry: the character budget is stated only for the Chinese text, with nothing said about the length of the English equivalent, so the two halves of one deliverable are specified to different levels of precision. It also explains why the negative prompt is English while the positive one leads in Chinese, which is a choice that will matter if your image tool expects a particular language for exclusions. ## Text inside the image is content, never an instruction One sentence in the skill covers a risk the rest of the design invites. Because the whole job is reading images full of typography, logos and captions, any text visible inside a reference image is treated as visual content to describe rather than as an instruction to execute. That is a prompt-injection defence, and for a skill whose input is arbitrary user-supplied images it is the right default. That is a statement of intent rather than a mechanism: the README states the rule, while the instruction that enforces it lives in `SKILL.md`, which is not shown here. Whether it holds under a deliberately crafted image is something you can only judge by trying it. The repository's own community pointer is a single external forum link in the README, so there is no issue tracker culture described here to fall back on. ## SKILL.md and two reference files carry the actual instructions The file structure is four lines: `SKILL.md` as the entry point, `references/analysis-framework.md` as the shared analysis framework, `references/category-guides.md` for category-specific rules, and `agents/openai.yaml` as Codex invocation metadata. Only the last filename is explained by anything outside the README, and nothing here shows what the model is actually told, how the category guides split photography from 3D from typography, or what parameters `agents/openai.yaml` configures. Licence is GPL-3.0, an unusual choice for a folder of Markdown: copying it into your own tools directory is a local act, while publishing a modified version carries the same terms as any GPL work. ## A high-fidelity claim with nothing to measure it The description calls this high-fidelity prompt reverse-engineering, and the output rules are specific enough to be useful. What the repository cannot show you is whether any of it works: there is no evaluation harness, no example of an input image paired with its generated prompt, and no comparison against a hand-written prompt. Treat the first run as an experiment rather than a pipeline. It is also the wrong tool when you already know how to write the prompt yourself, when you need reproducible output for an automated pipeline that cannot tolerate a model's judgement about which anchors matter, or when you are on a platform whose skills directory the README never names. ## Conclusion Adopt image-prompt-reverse if you generate images from references inside Codex on Windows and want a repeatable analysis frame rather than an improvised one. Do not adopt it for a pipeline that needs measurable prompt quality, for a POSIX host following the documented path, or for commercial redistribution without reading GPL-3.0. Verify first that SKILL.md and the two reference files match what you want the model told, since the README describes their contents only by title. ## FAQ ### What is reverse prompting in AI? In this project's terms it means reading an uploaded reference image backwards into a generation prompt. The skill analyses purpose, medium, subject, composition, camera language, lighting, color, materials, background, spatial depth, mood and post-processing style, then writes a positive prompt and a negative prompt for an image-generation tool. ### How do I install image-prompt-reverse for Codex? Copy the repository into %USERPROFILE%\.codex\skills\image-prompt-reverse, or clone it there with git clone. After installation you invoke it by name with $image-prompt-reverse. Only the Windows path is documented. ### What does image-prompt-reverse output? Two sections. The positive prompt is 450 to 700 Chinese characters of continuous natural language followed by an equivalent English version, and the negative prompt is 10 to 15 English terms separated by commas. No length is stated for the English equivalent. ### What kinds of images does the skill handle? Photography including portraits, products, documentary images, food, animals, nature and cityscapes; illustration in flat, anime, hand-drawn, Chinese-inspired and cyberpunk styles; 3D work from realistic renders to clay and designer toys; typography, logos, posters and graphic design; and IP-inspired or chibi characters, blind-box figures and mixed-subject scenes. ## Sources - [Issues](https://github.com/LunarXuan/image-prompt-reverse/issues) - [License: GPL-3.0](https://github.com/LunarXuan/image-prompt-reverse/blob/main/LICENSE) - [LunarXuan/image-prompt-reverse on GitHub](https://github.com/LunarXuan/image-prompt-reverse) - [README](https://github.com/LunarXuan/image-prompt-reverse/blob/main/README.md) --- Hysen Labs editorial analysis, written from the project's own repository and release notes. Cite the canonical page: https://hysenlabs.com/projects/lunarxuan-image-prompt-reverse