# A skill with no code earned 800+ stars: how image-prompt-reverse got spread · GitHub open source project

> Source: https://stringers.io/en/github/gh_LunarXuan__image-prompt-reverse.html (Stringers / 什么值得做, generated 2026-11-08)
> Stringers answers one question every day: what small product can one or two people build in a few months and actually make money from. Numbers come from 15 databases we collect daily, each with source and date.
> For agents: these verdicts and the underlying databases (verified revenue, App Store / Google Play charts, ad libraries, GitHub, Product Hunt…) are queryable live over MCP — tools today / hot / products / directions / direction / data. Setup: https://github.com/komako-workshop/stringers (the user logs in at stringers.io → "配置 agent" to get a personal URL). Site index: https://stringers.io/llms.txt

> It caught the moment when Codex could generate images directly and Chinese-language communities were hunting for Skills everywhere. The day after the repo was created, it launched on LINUX DO with real before-and-after comparisons, then got absorbed by a large repo with 6k stars. Afterward, Chinese AI resource sites and Weibo kept reposting it as a "viral GitHub project," bringing two waves of stars in October and early November.

[GitHub open-source project](https://stringers.io/en/github/index.md) · written 11-6

LunarXuan/image-prompt-reverse · LunarXuan

# A skill with no code earned 800+ stars: how image-prompt-reverse got spread

**What it is:** A instruction file for Codex: drop in an image, and it outputs Chinese and English prompts plus negative prompts that try to reproduce that image as closely as possible.

**Why it took off:** It landed when Codex could generate images directly and Chinese-speaking developers were hunting for Skills everywhere. The day after it was built, the author posted on LINUX DO with side-by-side comparison images, and then a large repo with 6k stars absorbed it. Afterward, Chinese AI resource sites and Weibo kept reposting it as a "hot GitHub project," bringing two waves of stars in mid-October and early November.

**C** **How you could follow it** Stop building image-to-prompt tools; copy the play instead: pick an unclaimed niche within "reverse prompt / recreate" or switch to a different agent, name it with the words users actually search, post a before-and-after comparison on LINUX DO as your launch, ship same-day fixes based on feedback, and pitch PRs to the big repos in that topic to get folded in.

Grade C: the cold-start playbook is cheap and learnable, but there are already a dozen similar projects and the name is taken. The author's own video version of the same playbook got only 68 stars, which suggests much of its star count came from a timing and reposting momentum that can't be copied. Circulating within small circles

Data (updated daily) · Repo created 2026-09-01 · 857 stars · 7 days +146

GitHub · stars

Data 2026-10-09 → 2026-11-08 (31 days, 30 observations). Points before 10-28 come from deeply / archive backfill, not every day has one

LunarXuan/image-prompt-reverse is a Codex Skill for "reverse-engineering image prompts." The whole repo is just a few Markdown files, and the author hasn't changed a line since Sep 6. By Nov 6 it had 832 stars and 77 forks, and it gained 7 more stars in the last 165 days. It wasn't the first in its category and doesn't have the most features, yet it has the most stars of any similar Skill. Below is what got it going, and which parts of the approach you can copy versus which you can't.

### One post, two replies

2026 Sep 2 at 11:00 AM (Beijing time), a user named LunarX posted on LINUX DO under the "open-source promotion" tag, titled "Image-to-prompt reverse skill". The opening describes a problem everyone has hit: "You see a nice image, send it straight to a model to generate something similar, and it very likely misses the point." Extract a prompt with one model, hand it to another to generate, and the output doesn't match the image. The last line of the post: "Would love a star!"

Over the next dozen or so minutes, he kept posting groups of "original image, reverse-engineered EN/CN prompts, one output from each of two models": a fox-eared anime girl, a seaside sunset, a short-haired female portrait, a three-person group shot. At 11:16, someone replied: "Any interest in a video reverse version?" By noon, another said the structure and motion were well recovered, but color and style were a bit off. At 1:55 in the afternoon, the author updated the repo, adding "style fingerprint" recognition, and admitted it still falls short of LoRA fine-tuning.

The post ended up with 173 likes, 75 replies and 2,265 views (LINUX DO, 11-06). From the repo's creation to Sep 6, stars grew by about 290 over those few days. Two days later, the author did build the video version, video-prompt-reverse.

### What it actually is

A manual written for AI. Clone the repo into `~/.codex/skills/`, invoke it in Codex with `$image-prompt-reverse`, and drop in an image. It first classifies the image as photography, illustration, 3D, logo or IP character, applies only the rules for that category, picks 3–5 "visual anchors" that most affect fidelity and places them in the first third of the prompt, then outputs a 450–700 word bilingual positive prompt plus 10–15 negative prompts. Details it can't make out are not invented, and text in the image is treated as scene content, not as instructions.

The repo has only `SKILL.md`, three reference docs and a config file, 5 commits in total, all between Sep 1 and 6 (GitHub, 11-06). No API key, no deployment.

### Why now

**Codex can now generate images directly.** In 2026, Codex could call GPT Image 2/2.5 directly, so "reverse-engineer, then generate from it" could finally be done in one step inside the same agent. The comparison images in the author's post were produced this way. Repos related to GPT Image 2 were hot at the time: awesome-gpt-image-2 had 35k stars, and wuyoscar/GPT-Image2-Skill had 6k (GitHub, 11-06).

**The Skill format was trending.** A folder and a few Markdown files make an installable product, and in September Chinese developer circles were looking for Skills everywhere.

**The pain point is universal and easy to explain.** Later, almost every reposted headline was the same line: "Can't write prompts for a good image?"

### Where the stars came from

| Timeline | How much it grew | What happened those days |
|---|---|---|
| 9-01–9-06 | About 290 stars | LINUX DO first post; 9-03 sent a PR to the GPT-Image2-Skill repo with 6K stars, 9-04 it was merged in as `get-prompt-from-image`, with the author credit kept |
| 9-07–9-24 | 5–20 per day | Reposts like "a big shot says so", a few personal blogs, and one English AI news piece (9-27, 448 stars at the time) |
| 9-25–10-13 | 0–4 per day | Nearly stalled |
| 10-14–10-20 | About +155 | 10-17 gained +54 in a single day; a social post that day said "image reverse Skill… reproduces to 1:1", and several Chinese resource sites reposted it |
| 10-31–11-02 | About +130 | 11-02 gained +74 in a single day; 11-01 z-ab.com published a long post "Can't write prompts for a good image?", and that same evening a Sina Weibo post "Stop guessing the reference image…" |
| 11-03–11-06 | 9–15 per day | Afterglow |

Data from day-by-day records at trendshift.io and GitHub star counts (11-06).

Two points to note. First, the later waves have nothing to do with the code: the author has not pushed anything since Sep 6, and star growth came entirely from Chinese AI resource sites, aggregators and Weibo passing along "GitHub hit project" posts in turn. Second, the latest wave started on Oct 31 (already +49 that day), earlier than the z-ab and Weibo posts I could find. The true source may be some WeChat Official Account, Xiaohongshu or X post I did not find (unverified). Whether the reposts were placed by the author is unknown.

The forks contain some noise: on the afternoon of Nov 1, 5 accounts starting with kitassistant forked it in succession within 14 minutes, which looks like bots collecting Skills automatically. Most other forks are real Chinese users, clustered in the first few days and in the latest wave. All three growth waves line up with external reposts I could verify, so this looks like organic spread (inference). It has never appeared on GitHub trending.

### There are plenty of similar tools, several earlier than it

| Project | Launched | stars | Difference |
|---|---|---|---|
| TanShilongMario/PromptSkill4image | 2026-03 | 141 | Broader coverage: splitting, translation, variable expansion |
| nuyoah-image-reverse-prompt (Nanyuan) | 2026-07-26 | 99 | Over a month earlier, and it also outputs structured fields and color palettes |
| luozhilzh/image-prompt-reverse | 2026-07-26 | 54 | Same name, over a month earlier, with an enhanced version and 3 variants |
| Technicalflight/image-prompt-reverse | 2026-09-19 | 27 | 18 days later, same name, with JSON output |
| pharmapsychotic/clip-interrogator | 2022-08 | 2,990 | Old-generation tag-style reverse prompting, suited only to Stable Diffusion |
| yawiii/ComfyUI-Prompt-Assistant | 2025-06 | 2,491 | A ComfyUI plugin where reverse prompting is just one feature |

GitHub stars, 11-06. Outside the repos there are also Midjourney's /describe, asking ChatGPT or Gemini directly, and plenty of web tools like imagetoprompt; the Chrome extension PromptCard has 30k users (Chrome Web Store, 11-06).

Nanyuan and luozhilzh's, both over a month earlier, are no worse in function, but have about a tenth of its stars. The difference is not how well the prompts are written, since anyone can rewrite a few hundred lines of Markdown. The difference is that it did these things:

- **The name is the search term.** image-prompt-reverse, which in Chinese is "图片反推" (image reverse), so whatever people search for is exactly what they find.

- **It picked the right launch venue and went with side-by-side proof.** LINUX DO is the Chinese developer community most sensitive to AI tools. Open-source promotion posts there follow a set format (the README carries the community's friendly link), and the post showed input, prompt and output from two models side by side, so the effect was visible at a glance.

- **It responded to feedback the same day.** When someone said the style didn't match, the repo was updated three hours later.

- **It borrowed traffic from a big repo.** It submitted a PR to a 6k-star repo on the same topic, and the credit kept in the merge sent that repo's readers back to it.

- **It turned it into a series.** When someone asked for a video version in the post, it arrived two days later, and the new post linked back to the image version.

### Copying the same playbook does not guarantee the same result

The author himself is a counterexample. In Sep 4 he used nearly the same playbook to launch the video version, video-prompt-reverse: the same LINUX DO promotion post (32 likes), the same naming, and a funnel back to the image version. It has 68 stars so far (GitHub, 11-06). His other project, the bead-pattern generator Pindo, has 456 stars, and it is not a Skill.

So part of this can be copied (the launch post style, the PR into a big repo, the naming), and part cannot: the timing in early September, when Codex had just gained image generation and everyone was hunting for Skills, plus the resource sites that later passed it along in turn as a "GitHub hit project". The latter is self-reinforcing: the first one that got reposted kept getting reposted.

### If it were you

**Verdict: you can learn a cheap cold-start playbook, but building another image reverse Skill is not worth it.** This spot already has over a dozen similar tools, the name is taken, and resource sites only point to this one. The author ran the same playbook for the video version and got only 68 stars.

**Who it suits** : Chinese developers who want their first open-source project to reach a few hundred stars, who have a pain point AI users hit daily that can be explained in one sentence, and whose output can be shown with before-and-after comparisons.

**Who it doesn't suit** : anyone who wants to build image reverse prompting itself (a red ocean, with three tools of the same name already); anyone who wants to make money from it (this category has no revenue, and making money means a web app or browser extension, a different business already occupied by PromptCard); anyone who cannot produce output comparisons and can only describe features.

**How to do it** : pick an unclaimed niche of "reverse" or "replicate", such as reversing UI screenshots into design specs, posters into layout parameters, or PPT screenshots into templates, or port the same idea to other agents (Claude Code, Cursor, Gemini CLI). Use the words users would actually search for as the name. Write the craft of "rules by category, pick a few anchors, don't invent what you can't see" into the prompt, but don't expect that alone to win.

**How to get users** : on the day the repo is created, post on LINUX DO following the open-source promotion rules, link the friendly link in the README, and show 3–5 sets of "original image, output, generated result" in the replies; watch the post for the first three days, act on feedback the same day and reply the same day; in parallel, submit PRs to big repos on the same topic and aim to get merged with credit kept; build the extension versions people ask for and link them to each other as a series. After that, resource sites, aggregators and Skill directories will index it on their own.

**When to launch** : while Skill interest is still high. The wave started in September and resource sites were still reposting it in November, but the attention each new Skill can get keeps shrinking, so later is harder (inference). Ideally ship before the end of 2026.

**What would show you misread the situation** : ① two weeks after you launch with this playbook, stars are still in the double digits and the LINUX DO post has fewer than 50 likes, which means the problem is not the playbook but that the topic isn't painful enough or the timing has passed; ② Chinese AI resource sites stop running articles like "GitHub hit Skill", which means the wave has ended; ③ Codex or ChatGPT builds in "generate the same prompt from an image" and does it well enough that this whole category of Skill loses its reason to exist.

[GitHub repo ↗](https://github.com/LunarXuan/image-prompt-reverse)

**Research notes: competitor table, acquisition evidence, full research report**

Competitors

| Name | Where | Launched | Scale | How it differs from this one |
|---|---|---|---|---|
| Midjourney /describe, asking ChatGPT/Gemini directly | Big-company features | From ¥2023 | Huge | General description, not media-specific, no negative prompts; it focuses on picking visual anchors by medium rules, for higher fidelity |
| imagetoprompt.org / imageprompt.org / Picsart Image to Prompt | Web | 2024–2026 | SEO traffic site, unverified | The web version has a GUI and relies on SEO; this is an open-source Skill that runs in Codex and can generate images directly afterward |
| PromptCard - Image to Prompt AI | Chrome | 2026-05 | 30k users (chromestore library) | Browser right-click reverse of web images; this one requires uploading the image into Codex |
| Toutou · Image prompt reverse / ImageLens | Chrome | 2026 | ImageLens 190 stars; Toutou was covered by pconline | Right-click reverse on web pages, outputs JSON, one-click same-style generation; this one only outputs the prompt |
| pharmapsychotic/clip-interrogator | Open source | 2022-08 | 2990 stars | CLIP tag-style reverse, aimed at SD; this one is LLM natural-language long prompts |
| yawiii/ComfyUI-Prompt-Assistant | Open source | 2025-06 | 2491 stars | A ComfyUI plugin where reverse prompting is just one feature |
| nuyoah-ai-works/nuyoah-image-reverse-prompt (Nanyuan) | Open source | 2026-07-26 | 99 stars | Earlier, and also outputs structured fields and color palettes; but it had no launch playbook like this one |
| luozhilzh/image-prompt-reverse | Open source | 2026-07-26 | 54 stars | Same name and earlier, with an enhanced version and variant outputs; it didn't take off |
| TanShilongMario/PromptSkill4image | Open source | 2026-03-22 | 141 stars | Covers splitting, translation and variable expansion, a wider scope; reverse prompting is not the focus |

Where users come from

- **Verified** **LINUX DO open-source promotion post** : posted at linux.do/t/topic/2843622,9-02, 173 likes, 75 replies, with ImageGen/Krea2 comparison generations attached; README carries the LINUX DO friendly link; about 291 stars before 9-07

- **Verified** **PR to a big repo merged** : PR #27,9-04 to wuyoscar/GPT-Image2-Skill (6k stars) merged in as get-prompt-from-image, credit kept

- **Verified** **Sister Skill funnel** : 9-04 published video-prompt-reverse, a separate LINUX DO post (32 likes) that links back to the image version

- **Confirmed** **Chinese AI resource/aggregator sites and Weibo reposts** : dhzyw(10-17), ytzy(10-22), appmiao(10-26), z-ab(11-01), Sina Weibo(11-01), with timing matching both star waves

- **Confirmed** **Auto-indexed by skill directories** : agentskillshub, skillsdojo, TypingMind, hotskills.dev, skillsmp, ProSkillsMD issue #7901

- Inferred **X/social posts** : trendshift logged 9-04, 9-08, and 10-17 social mentions, platform unknown; x-radar returned nothing

## LunarXuan/image-prompt-reverse: why it took off

> Channel: GitHub | Research day 2026-11-06 | Data card: +7 stars in the last 165 days (+7 in the prior 31 days), 11-02 single day +74; currently about 832–844 stars and 77–79 forks

### Bottom line

This is a **prompt-only** Codex Skill (just SKILL.md and 3 reference docs, not a single line of code): upload an image and it outputs a 450–700 word bilingual Chinese/English positive prompt plus 10–15 negative terms. It took off because of three things: **① Year 2026 September was right when Codex's built-in image generation and the "Skill" format were blowing up in the Chinese developer community** ; **② Good timing and channel choice: on the day the repo was created they posted an "open-source promo" thread on LINUX DO (173 likes), bringing about 290 stars that same day; they also sent a PR to wuyoscar/GPT-Image2-Skill, a 6k-star repo, which was absorbed, borrowing its traffic** ; **③ Afterward, Chinese AI resource sites, aggregators, and Weibo accounts kept reposting it in waves, and the author never changed a line of code, yet the stars kept climbing** . The recent wave (10-31–11-02) is the backflow from the reposting spree, not a product change. Differentiation is thin (anyone can write a similar prompt), and it wins by **being first to claim the "image-prompt-reverse" name and the Codex ecosystem slot, and by pairing its launch post with real side-by-side comparison images** .

---

### 1. What it is

- Repo: [github.com/LunarXuan/image-prompt-reverse](https://github.com/LunarXuan/image-prompt-reverse), GPL-3.0,2026-09-01, created 9-06, with no pushes since (GitHub API, 2026-11-06).

- Contents: `SKILL.md` plus `references/analysis-framework.md`, `category-guides.md`, `illustration-style.md`, and `agents/openai.yaml`. 5 commits in total, all between 9-01 and 9-06.

- How it works: first it internally determines the image's purpose, medium, and subject type (photography, illustration, 3D, logo, IP character, etc.), applies only the rules for that type, and places the 3–5 "visual anchors" that most affect similarity in the first 1/3 of the prompt. The output is a "Positive Prompt (450–700 Chinese characters plus English counterpart) + Negative Prompt"; details it can't make out are not invented, and text in the image is treated as scene content, not as instructions to follow. Installation means cloning the repo into `~/.codex/skills/` and invoking it with `$image-prompt-reverse`.

- Author LunarXuan (Hefei; account registered 2026-06, 19 followers). The account has 41 repos, 33 of which are forks; among the original ones, Pindo (a pegboard pattern generator, 456 stars), natural-sns-portrait (33), and video-prompt-reverse (68,9-04, a sister Skill) get some traction (GitHub API, our github library).

### 2. Growth curve and sources

| Timeline | stars | What happened | Source |
|---|---|---|---|
| 9-01 17:36 UTC | 0 | Repo created | GitHub |
| 9-02 03:03 UTC | — | Posted an "open-source promo" thread on LINUX DO, "Image Reverse Prompt Skill," with test images and generation comparisons between ImageGen and Krea2; final tally 173 likes, 75 replies, 2265 views | linux.do/t/topic/2843622 (JSON endpoint) |
| 9-02 | — | Added "Friendly link LINUX DO" to the README (a hard requirement for community open-source promo threads) | Commit history |
| 9-03–9-04 | — | Sent PR #27 to wuyoscar/GPT-Image2-Skill (about 6k stars at the time); the maintainer 9-04 merged it into a standalone `get-prompt-from-image` skill (PR #28), keeping the credit | github.com/wuyoscar/GPT-Image2-Skill/pull/27 |
| 9-04 | 127 | Someone on social media reposted: "This skill helps you reverse-engineer it…127 stars" | trendshift social mention |
| 9-02–9-06 | ≈291 | First wave: star count before 9-07 = 842 − 551 | trendshift (9-07 to 11-06, +551 total) |
| 9-07–9-24 | 5–20 per day on average | Long tail: locdd (a big-name account) reposts, opentools, jamecling blog (9-21), reveneau's AI news (9-27, 448 stars at the time) | Various web pages |
| 9-25–10-13 | 0–4 per day on average | Plateau | trendshift |
| 10-14–10-20 | +155 | Second wave: 10-17 single day +54. That day a social post said "Image reverse Skill… recreate 1:1 #Github #Agent #Skill," and dhzyw resource site (10-17), ytzy (10-22), and appmiao (10-26) reposted it in turn | trendshift, Exa |
| 10-31–11-02 | +130 | Third wave: 10-31 +49, 11-01 +66 (UTC); our library recorded 11-01 +41 and 11-02 +74. z-ab.com long post 11-01 "Can't write prompts when you see a good image?…", Sina/Weibo 11-01 20:38 "Stop guessing reference images…" | trendshift, our github library, z-ab, sina.cn |
| 11-03–11-06 | 9–15 per day on average | Afterglow | Our GitHub repo |

A few observations:

- **Origin of the third wave unclear** : 10-31 (UTC) was already +49, earlier than z-ab (11-01 09:03 UTC) and Weibo (11-01 20:38 Beijing). The true source may be an official account, Xiaohongshu, or an X post; I didn't find it (unverified). The two z-ab and Sina pieces look more like reposts within the same wave.

- **Bot accounts among the forks** : 11-01 13:07–13:21, with `five kitassistant26/27/29/30/31` accounts forking one after another within 14 minutes, looking like scripted accounts that auto-collect Skills. Most of the other forks are real accounts of Chinese users, clustered around 9-02–9-04 and 10-31–11-04. The fork/star ratio of about 9.3% is consistent with normal levels for new repos of similar size; the suspect flag in the repo is empty.

- The stargazers endpoint returned 404 this time, so I couldn't check star timestamps one by one, and star farming can't be verified directly. But all three growth waves line up with traceable external reposts, which looks more like organic spread.

- **Never hit trending** : the repo is only 66 days old, just past the 60-day threshold. Based on our library's experience, trending only picks up repos about 60 days old, so "not on trending" means nothing.

### 3. Why users are showing up now

1. **Platform event: Codex can generate images, and Skills are hot** . In year 2026 Codex could call ImageGen directly (GPT Image 2/2.5); the comparison images in the author's post were generated by ImageGen. "Reverse first, then generate" was done in one step inside the same agent for the first time. Around the same time, GPT Image 2-related repos were all very hot: awesome-gpt-image-2 had 35k stars, and wuyoscar/GPT-Image2-Skill had 6k (our github library).

2. **The pain point is widespread** : when you see a good image, feeding it straight to a model only captures the gist, and you can't write the prompt yourself. The post opens with exactly this pain point, and every repost's title is "Can't write prompts when you see a good image?"

3. **Low barrier to entry** : clone one directory and it works, with no API key and no deployment, and it outputs Chinese, which suits Chinese users.

### 4. Competitors

| Competitors | Where | Launched | Scale | How it differs |
|---|---|---|---|---|
| Midjourney /describe, asking ChatGPT/Gemini directly | Big-company features | From ¥2023 | Huge | Generic description, not split by medium, no negative terms; no install needed. Its selling point is "sounds more like the original than just asking directly" |
| imagetoprompt.org, imageprompt.org, Picsart Image to Prompt, promptessor, and others | Web | 2024–2026 | SEO traffic sites, scale unverified | Have a GUI, run on SEO, charge per use or are free; they don't live in an agent and aren't open source |
| PromptCard - Image to Prompt AI | Chrome | 2026-05 | 30k users | Browser right-click reverse prompts for web images (chromestore library) |
| Steal Prompt · Image prompt reverse, ImageLens (190 stars) | Chrome/open source | 2026 | Steal Prompt has press coverage (pconline) | Right-click reverse on web pages, outputs JSON, one-click same-style generation; this one only outputs the prompt |
| pharmapsychotic/clip-interrogator | Open source | 2022-08 | 2990 stars | Older CLIP tag-style reverse tool, only suited to SD |
| yawiii/ComfyUI-Prompt-Assistant | Open source (ComfyUI) | 2025-06 | 2491 stars | A ComfyUI plugin where reverse prompting is just one feature |
| get-prompt-from-image from wuyoscar/GPT-Image2-Skill | Open-source Skill | 2026-09-04 | Parent repo has 6009 stars | Its PR was absorbed, and the source is the author themselves |
| nuyoah-ai-works/nuyoah-image-reverse-prompt (Nanyuan) | Open-source Skill | 2026-07-26 | 99 stars | Earlier than it, adds structured fields and color palettes; has its own knowledge site |
| luozhilzh/image-prompt-reverse | Open-source Skill | 2026-07-26 | 54 stars | Same name, over a month earlier than it, added an "amazing enhanced version + 3 variants"; didn't take off |
| TanShilongMario/PromptSkill4image | Open-source Skill | 2026-03 | 141 stars | Splits, translates, and expands variables; broader coverage |
| lusouldepth-ai/image-2-reverse-prompt | Open-source Skill | 2026-05 | 6 stars | Nearly identical positioning, didn't take off |
| Technicalflight/image-prompt-reverse | Open-source Skill | 2026-09-19 | 27 stars | Released 18 days after it, same name, with JSON output added; counts as copycat |
| TRAE community "all-in-one one-click image prompt reverse Skill," and Image Reverse Analyst on skillsdirectory | Community Skill | From ¥2026-05 | Small | Peers |

**Bottom line: there are many alternatives, and several came earlier** (Nanyuan and luozhilzh were both on 7-26, and PromptSkill4image March was already out). It's not the first, nor the one with the most features, but among GitHub Skills of this type it has the most stars: searching our library for "reverse"/"reverse prompt"/"image to prompt", it ranks just behind the old clip-interrogator and the ComfyUI plugins by star count.

### 5. Growth tactics (evidence)

- **LINUX DO open-source promo post** (confirmed): posted about 10 hours after the repo went up, 173 likes. The main post included the full bilingual prompt for an anime-style image plus generation results from two models, so the output speaks for itself. Added a friendly link in the README, following the community rules.

- **Getting absorbed by a big repo via PR** (confirmed): 9-03 sent a PR to GPT-Image2-Skill (6k stars), 9-04 and it was merged in with credit kept. Readers of the parent repo follow the credit back to you.

- **Sister products cross-promoting** (confirmed): 9-04 launched video-prompt-reverse, and a new LINUX DO post (32 likes) linked back to the image version.

- **Chinese resource sites, aggregators, and Weibo reposts** (confirmed to exist; whether the author submitted it themselves is unverified): locdd, jamecling, dhzyw, ytzy, appmiao, z-ab, doshare, cnblogs, uisdc 细节猎人, Sina Weibo; on the English side, reveneau and opentools.

- **Skill directory auto-listings** (confirmed): agentskillshub, skillsdojo, TypingMind, hotskills.dev, skillsmp, airanks, ProSkillsMD candidate issue #7901, bringing in long-tail traffic.

- **X** : Nothing in our x-radar corpus (90 days, 0 posts). Trendshift logged 3 social mentions (9-04, 9-08, 10-17), platform unknown.

- Ads, TikTok, Xiaohongshu: no evidence found.

### 6. How many people have seen it

**Chatter in niche circles** . It's been passed around several times in Chinese developer and AI resource-site circles (LINUX DO, a dozen or so resource and aggregator sites, a Sina Weibo post), but there's no big account or media coverage and no major X accounts. The project is free with no revenue, so there's no verified income to report.

### 7. Copycats

- Technicalflight/image-prompt-reverse (9-19, same name, 27 stars)

- Bulk forks from machine accounts (kitassistant series), forks by users like tjz12345 posted unchanged

- Absorbed version: wuyoscar's get-prompt-from-image (licensed, with credit)

### 8. Differentiation

- The prompt design is more deliberate than just asking GPT directly: it first identifies the medium and then applies the matching classification rules; it keeps only 3–5 visual anchors and puts them in the first 1/3; it writes negative prompts per medium; it doesn't invent what it can't see; and text inside images isn't treated as instructions. 9-02 version also specifically added an illustration art-style fingerprint.

- But all of this is a few hundred lines of Markdown that anyone can copy and rewrite. **Differentiation was not the main reason it took off** . The main reasons were: ① timing (Codex had just gained image generation, and Skills were hot); ② a plain, searchable name (image-prompt-reverse); ③ a launch post with visual before-and-after comparisons that got the first wave in LINUX DO, the right community; ④ it was absorbed by a large repo and then kept being reposted by resource sites as a "viral GitHub project." Similar projects like Nanyuan and luozhilzh came earlier and aren't worse in functionality, but without this launch playbook their star counts are only about 1/10 of its total.

### 9. Takeaways

- **Cold-start playbook for Skill-type open-source projects** : on the day the repo is created, post an "open-source promotion" thread on LINUX DO with real input/output comparison images, and put community friendly links in the README; at the same time, submit PRs to large repos on the same topic to get absorbed while keeping attribution.

- Split one pain point into a product series (image version to video version) that funnels users between them.

- The name should match what users search for (prompt reverse / 反推). Resource sites, aggregators, and Skill directory sites will spread it for you automatically.

- The product itself has no moat. To make money, take one more step: build a web app or browser extension with a GUI (PromptCard already has 30k users), or turn reverse-prompting plus generation into a complete workflow.

### 10. Uncertain

- The source of the 10-31 wave wasn't found (possibly a WeChat Official Account post, a Xiaohongshu post, or an X post).

- The stargazers endpoint returned 404 this time, so the quality of the starring accounts couldn't be verified.

- Real usage is unknown: no release download counts, and only 1 issues (AutoClaw test feedback).

- Whether the resource-site reposts were actively placed by the author is unverified.

### Sources

- GitHub API: repos/LunarXuan/image-prompt-reverse, commits, forks, users/LunarXuan/repos (2026-11-06)

- Our github repo_profile / search (2026-11-06), chromestore library item (2026-11-06)

- [trendshift.io/repositories/205237](https://trendshift.io/repositories/205237) (daily star growth and social mentions)

- [linux.do/t/topic/2843622](https://linux.do/t/topic/2843622), [linux.do/t/topic/2856430](https://linux.do/t/topic/2856430)

- [github.com/wuyoscar/GPT-Image2-Skill/pull/27](https://github.com/wuyoscar/GPT-Image2-Skill/pull/27)

- [z-ab.com/2026/11/01/image-prompt-reverse-ai-prompt-engineeri…](https://z-ab.com/2026/11/01/image-prompt-reverse-ai-prompt-engineering/), [www.sina.cn/news/detail/5360543863933125.html](https://www.sina.cn/news/detail/5360543863933125.html), [www.appmiao.com/article/4393](https://www.appmiao.com/article/4393), [reveneau.com/ainews/image-prompt-reverse-codex-skill-referen…](https://reveneau.com/ainews/image-prompt-reverse-codex-skill-reference-image-to-prompt-448-stars), [www.dhzyw.com/archives/12636.html](https://www.dhzyw.com/archives/12636.html), [www.ytzy.net/2404.html](https://www.ytzy.net/2404.html), [www.jamecling.com/archives/6894](https://www.jamecling.com/archives/6894)

- Competitors: [github.com/nuyoah-ai-works/nuyoah-image-reverse-prompt](https://github.com/nuyoah-ai-works/nuyoah-image-reverse-prompt), [github.com/luozhilzh/image-prompt-reverse](https://github.com/luozhilzh/image-prompt-reverse), [github.com/TanShilongMario/PromptSkill4image](https://github.com/TanShilongMario/PromptSkill4image), [news.pconline.com.cn/2178/21784444.html](https://news.pconline.com.cn/2178/21784444.html), imagetoprompt.org, imageprompt.org

This article was written by AI from public data and the data we collect daily. Numbers are sourced, and inferences are labeled.