image-prompt-reverse: Reverse AI-Generated Images Back into Bilingual Chinese/English Prompts

Who It's For

• Anyone who wants to know "what prompt was originally used" for an AI-generated image: hand the image to a Skills-capable agent, trigger "reverse prompt" / "image to prompt", and get bilingual (Chinese/English) reconstructed prompts, style tags, and a structured JSON prompt.
• Users who already have an agent skill directory and are willing to install manually. The project explicitly states you only need to copy the entire directory into the agent's skill directory (e.g., WorkBuddy's ~/.workbuddy/skills/).
• Teams that need machine-verifiable output: the project ships its own validation script scripts/validate_analysis.py, which depends only on the Python 3.8+ standard library — zero third-party dependencies.
• Not suitable when: ⚠️ This project is licensed under AGPL-3.0. Any commercial use (including integration into commercial products, commercial services, SaaS, paid tools, or commercial operations based on it) requires prior written authorization from the author. Personal learning and non-commercial analysis are fine to try.
• The evidence does not state the project's requirements regarding image formats, model backends, or specific agent versions; these prerequisites are unknown, so refer to the repository's latest documentation.

First Try in 3 Minutes

1. Prerequisites: You can access the GitHub repository; Python 3.8+ is available locally (only for the optional validation script); your agent supports Skills and has a skill directory (officially, WorkBuddy's ~/.workbuddy/skills/ is given as the example).
2. Install: After downloading the repository, copy the entire image-prompt-reverse directory into the agent's skill directory, e.g., ~/.workbuddy/skills/. This is the only installation method given by the official documentation; skill directory paths for other platforms are not listed officially and are unknown.
3. Trigger: In the agent conversation, send an image and say "reverse prompt" (反推提示词) or "image to prompt" (图片转提示词). These are the officially documented trigger phrases.
4. Success signals: After viewing the image, the agent produces three things — a bilingual (Chinese/English) reconstructed prompt, style tags, and a structured JSON prompt. These are text outputs; the official documentation does not state whether the agent writes files, which is unknown.
5. Optional validation: Run the following on the resulting analysis.json:

python scripts/validate_analysis.py analysis.json

Exit code 0 means it passes; 1 means the language buckets are mixed or the tags fail; 2 means the schema is missing fields.

Note: The validation command reads and writes the JSON file you specify. Whether the script has network access or other side effects beyond file read/write is not stated in the evidence and is unknown. The first attempt itself only has the agent perform analysis and reconstruction — it does not involve modifying your other files — but the agent's image reading and output behavior depends on your actual client.

Use Cases

Case 1: Completing prompt records for collected AI images. Scenario: You're building a personal asset library and want to add "what the original prompt roughly was" to a generated image. Send the image to the agent and say "image to prompt"; the agent executes its core workflow: a forensic-style reconstruction of the most likely original generation prompt. Expected output: a Chinese prompt, an English prompt, a list of style tags, and a structured JSON prompt. Run the JSON through python scripts/validate_analysis.py analysis.json; an exit code of 0 means the output structure is valid.

Case 2: Comparing style tags across multiple images. Scenario: You want to summarize the common style of a group of similar generated images. Send images one by one, triggering analysis with "reverse prompt", and manually compare the style tags and bilingual prompts from each output. The evidence confirms only that "each image gets an analysis containing style tags and a JSON prompt"; whether the project offers batch processing or comparison features is not stated officially — do not assume you can upload multiple images at once for automatic summarization.

Usage Boundaries

• Installation: The official documentation provides only one method: "copy the entire directory into the agent's skill directory". There is no evidence of a Skills CLI, plugin marketplace, or other installation entry points; do not assume the same paths work on clients that are not listed.
• Triggers and output: Known trigger phrases are "reverse prompt" (反推提示词) and "image to prompt" (图片转提示词); outputs are bilingual prompts, style tags, and a structured JSON prompt. The reverse-engineered result is a reconstruction of "the most likely original prompt" and is not guaranteed to equal the actual original prompt.
• Commands and side effects: The only official command is the validation script, with zero third-party dependencies (Python 3.8+ standard library), reading and writing the JSON file you specify. The script's full read/write scope and network behavior are not stated in the evidence and are unknown.
• License restrictions: AGPL-3.0 — any commercial use requires prior written authorization from the author. This is a hard prerequisite you must confirm before use.
• Unconfirmed information: Supported image formats, whether output language is configurable, directory paths for non-WorkBuddy clients, and whether local files will be written or modified are all not stated in the official evidence; refer to the repository's latest README.