Getting it into your agent
One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.
npx agentmods add skills/jing-hy/picturereader-zcode/image-readingnpx skills add jing-hy/picturereader-zcode --skill image-readinggit clone --depth 1 https://github.com/jing-hy/picturereader-zcodeWrote this? Show the measurements
A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.
[](https://agentmods.dev/skills/jing-hy/picturereader-zcode/image-reading)<a href="https://agentmods.dev/skills/jing-hy/picturereader-zcode/image-reading"><img src="https://agentmods.dev/badge/skills/jing-hy/picturereader-zcode/image-reading.svg" alt="Measured on agentmods" height="20"></a>What it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5 | $0.00072 | $0.02007 |
| Opus 5 | $0.00036 | $0.01004 |
| Sonnet 5 | $0.00014 | $0.00401 |
| Haiku 4.5 | $0.00007 | $0.00201 |
Grade A, and why
image-reading scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 3d ago.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Nothing flagged
None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.
How it starts
The opening of the file, as written. The whole thing — 97 lines — stays where its author put it; the contents beside it link to each section on GitHub.
读图方法论(image-reading)
目标:像多模态模型一样"看"图并输出连贯描述,每个结论可追溯、可验证。 本 skill 由 experience / skill / principle / insight 四层知识构成(按 Gogomoe 知识框架分类,教训均来自对真实图片的实测复盘)。
操作流程(skill)
1. 全局定调(第一轮扫描)
用默认参数(size=40)全图扫描,读四个字段:
hue families(最高优先级):按纯色相分族的真实占比。暗调/低饱和场景的 真实颜色只在这里——colors by area灰白占比高不代表画面灰白。structure:平行条带/对称性(解读见 insights)。texture:rough 高=写实照片;smooth 高=扁平或水面/天空/雾(见 insights)。regions:大结构的位置/大小/颜色。
2. 找主体(全局→局部,主动验证)
- 对颜色异常区、深色大块、相邻竖长色块、小色块密集区用
px_per_cell定向放大 (值越小越细:8-12 看轮廓,4-6 看结构,2-3 看细节;区域不够小时工具会提示实际密度,缩小 focus/region 重试)。 - 放大后按形状解读:头+肩+躯干=人物;弧线+对称明暗=圆柱/球/装置; 竖直细长结构=石柱/塔/杆;交替细条=面板/栅格。
- 主体可能与背景低对比而"隐形"(见 insights 4)——怀疑处必须放大确认,不能因 regions 未单列就跳过。
3. 文字验证
- 疑似文字/标识/UI →
image_ocr(region/focus 限定)。 - Windows 引擎读不出但怀疑有字 →
engine="paddle"重试(发光/弯曲/游戏渲染文字)。 - OCR 结果优先于模型描述(见 insights 3)。
4. 材质判断
image_sample 对小块区域 8×8 取样,看 RGB 分布与 contrast 统计
(平滑渐变=天空/皮肤/水面;高对比条纹=金属/木纹;暗绿 G>R>B=植物/涂装)。
5. 综合描述
输出连贯描述(场景/主体/环境光线/细节),每个结论标注证据等级: 实锤(有像素/OCR/取样数据)vs 推断(基于结构推测,用"看起来像")。 优先引用具体数字;不确定就说不确定,绝不编造。
行为准则(principles)
- 证据分级:任何结论标注"实测"或"推断";推断必须说明依据。
- 数字优先:用具体指标("蓝色调 74%""对称 80%""OCR 读出 1.00")支撑描述,不用模糊形容词代替。
- 先全局后局部:第一轮定调,第二轮定向放大验证,不跳步。
- 怀疑即验证:对任何"可能漏掉的主体",用放大/取样/OCR 验证后再下结论。
- 不编造:不确定就说明;模型(含多模态)的描述不可直接当作事实(见 insights 3)。
规律性洞察(insights,跨图归纳)
- 暗调场景的真实颜色只在 hue families 里:低饱和/暗色调(暮色、雾中、夜景)
会被 14 色色板压成灰黑,
colors的灰白占比是假象——hue families 按纯色相分族不受影响。 - 高对称 ≠ 一定人造物:水面倒影/镜像构图也高度对称。区分看:平滑大面积 (水面/天空 smooth 高)+ 水天分界线(上亮下暗、上下镜像)+ 竖直细长结构(石柱) = 湖泊/自然镜像;纹理复杂、颜色单调、几何硬边 = 人造建筑/装置。
- 小模型读小字不可靠:多模态小模型对低分辨率文字会幻觉(全图"读出"内容、 裁剪后承认没有);发光/弯曲/艺术字 Windows OCR 也失效——文字一律以 OCR 实读为准。
- 低对比主体"隐形":暗色物体(如深色服装人物)在暗背景中融入背景黑块, 粗网格和 regions 都不会标出——对深色区域主动放大是唯一可靠发现方式。
- 平滑大面积 ≠ 扁平简笔画:水面、天空、雾气、墙面都平滑(smooth 高), 需结合色调/结构/场景判断,不能仅凭 smooth 判定"扁平"。
- "像什么"和"是什么"要分开:结构证据(对称/形状/色调)支撑"像什么"; "是什么"需要 OCR/取样/更强证据,不满足时保持推断。
- hue families 是场景类型指纹(34 张图训练归纳):
- cyan 高(>60%)= 水/雾/湖泊/晨雾场景(东方水景、浓雾遗址)
- green 高(>40%)= 森林/竹林/草地/苔藓
- orange 或 red 高 = 红披风/暖色服饰人物、火光、晚霞
- blue 高(>70%)= 夜晚/冷色科幻场景
- achromatic 高 + rough 高 = 废墟/岩石/暗环境
- green + yellow 双高 = 翠绿能量带/发光植被/浮空仙境
- 对称高 + 中央竖直结构 = 中央主体(瀑布/树/大门)居中构图
- 多模态模型的颜色描述对"发光/能量"不可靠(训练中反复出现):把实测为 cyan/blue/green 的冷色发光(屏幕光、能量屏障、雾中光柱)系统性说成"粉红/紫色"。 发光元素的颜色一律以 hue 实测为准。
- 人物识别信号:orange/red 主调 + 局部暖色小块 + 对称 = 人物服饰候选; 游戏角色常穿红/橙(红披风、红发、暖色战斗服),识别到暖色主调时应主动放大找人物。
- 品牌/游戏名/标题文字:多模态模型会猜错("原神""崩坏3"实际是明日方舟终末地), 必须 PaddleOCR 实读(游戏 HUD 底部常带游戏名/参数/水印)。
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- 3d ago First seen · 97 lines · 72 tokens per session scan A 24ab81cd0249
image-reading is a skill published in the GitHub repository jing-hy/picturereader-zcode (2 stars, last pushed 17d ago), licensed MIT. It adds 72 tokens to every session and 2,007 once invoked, about $0.0004 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-31.
Other skills, from other repositories
vision
Nested swiss-knife reference for image understanding — a decision tree that routes between three paths depending on what the agent has access to: (1) the built-in vision tool if the LLM provider supports image input, (2) the sibling minimax-cli reference if a usable MiniMax preset/key slot is available, or (3) a local…
paper-image-extractor
Extract figures from papers — prioritizes arXiv source package for high-quality images.
glmv-stock-analyst
股票分析与涨跌预测分析。 在用户表达分析、判断或预测意图时触发,如“分析一下腾讯”、“0700最近走势如何”、“XX能不能买”、“预测一下后续走势”、“生成一份分析报告”等; 支持港股、A股、美股,整合多源数据(包括新闻、基本面、技术面、资金流及宏观信息)进行多维综合分析,输出图文结合、包含可视化图表的结构化分析报告。 对于简单查询类需求(如“腾讯当前价格是多少”、“茅台代码是什么”)不触发本skill, 直接通过websearch 能力搜索并总结。 ⚠️ 需要多模态主模型支持(如 glm-5v-turbo),主模型需能读取图片。.
glmv-web-replication
Frontend visual replication skill. Explores a target website’s publicly visible pages via Playwright MCP or agent-browser, captures screenshots and layout information, then generates a static or client-side frontend replica that approximates the original’s visual appearance and page structure. This skill replicates…
glm-image-gen
Official skill for generating high-quality images from text prompts using ZhiPu GLM-Image API. Excellent at scientific illustrations, high-quality portraits, social media graphics, and commercial posters. Supports multiple aspect ratios, HD quality, and watermark control. Use this skill when the user wants to generate…
glmv-pdf-to-ppt
Convert a PDF (research paper, report, or any document) into a polished multi-slide HTML presentation with a structured outline JSON and summary markdown. Trigger this skill when the user mentions making slides or a PPT from a PDF — in Chinese or English.