@yuhan1124/draw-prompt 0.4.0

package/references/gallery.md

@@ -0,0 +1,188 @@
+# gallery.md —— 自带金标准范例库（baseline，所有用户共享）
+
+这是 skill **自带**的高质量 prompt 范例库,不依赖任何外部 skill。每条都是按
+`compile.md` 七段骨架精构的「金标准」,覆盖主流品类。编译时先在这里找最近的 1-2 条
+对标改编,产出的 baseline 就已经很好——**个人偏好(profile/samples)只是在此之上叠的
+delta,不是从零开始**。
+
+> 本机若装了 `gpt-image` skill,可把它 `references/` 下的 162 条 gallery 当**可选扩展**
+> 额外检索(注意其外部作者归属);但本库已能独立支撑高质量出图。
+>
+> 全部为原创金标准,可自由使用/分发。每条标注:推荐 `size`·`quality` + 它示范的法则。
+
+## 路由索引
+
+| # | 品类 | 关键词 |
+|---|---|---|
+| 1 | 中文促销海报 | poster, 海报, 促销, 中文排版 |
+| 2 | 极简英文活动海报 | poster, event, swiss, typography |
+| 3 | 移动 App UI | ui, mockup, app, 界面 |
+| 4 | Web 仪表盘 UI | ui, dashboard, saas, web |
+| 5 | 科普信息图 | infographic, 科普, 图解, 固定区域 |
+| 6 | 学术/系统架构图 | diagram, architecture, 论文图, 数据图 |
+| 7 | 产品渲染(JSON 式) | product, 产品, 渲染, 食品, 电商 |
+| 8 | 风格化插画 | illustration, 插画, 风格化 |
+| 9 | 写实摄影 | photography, 写实, 街拍, 场景 |
+| 10 | 角色设计参考单 | character, 角色, 立绘, 多视图 |
+| 11 | 等距 3D 场景 | isometric, 3d, 等距, 微缩 |
+| 12 | Logo / 品牌系统 | logo, brand, 品牌, 标识 |
+
+---
+
+### 1 · 中文促销海报
+- 推荐:`portrait` (1024x1536) · `high`
+- 示范:画幅优先 · 精确文字全加引号 · 促销层级三瞥测试 · 否定淘宝味
+
+```text
+Design a 3:4 vertical promotional poster. Restrained, light-luxury commercial style — clean grid, generous negative space, sophisticated not cheap; not an e-commerce template.
+
+Main subject: [一句具体主体,如 a cold-brew tea cup with citrus and ice]. Soft studio light, shallow depth of field.
+
+The poster must accurately display the following exact Simplified Chinese copy, in a clear size hierarchy:
+Headline (largest): "[主标题]"
+Sub-slogan: "[一句主张]"
+Price / discount block (prominent, highly legible): "[价格1]" / "[价格2]" / "[优惠]"
+Info modules: "[模块A]" / "[模块B]"
+Footer fine print: "[活动时间]" / "图片仅供参考,以实际为准"
+Brand mark: "[BRAND]"
+
+Keep numbers, prices and Chinese typography crisp, legible, aligned; promotional hierarchy readable from a distance.
+Avoid: Taobao-style red/yellow clutter, fake brand logos, garbled or wrong Chinese characters, high-saturation gradients.
+```
+
+### 2 · 极简英文活动海报
+- 推荐:`portrait` (1024x1536) · `high`
+- 示范:风格锚点具体有边界(Swiss grid) · 文字即主体 · 留白
+
+```text
+Design a 3:4 vertical event poster in a Swiss International Typographic style — strict grid, Helvetica-like sans-serif, bold asymmetric layout, one accent color on off-white, lots of negative space.
+
+Exact text, in clear hierarchy:
+Title (very large): "[TITLE]"
+Date / venue: "[DATE]" · "[VENUE]"
+Supporting line: "[one line]"
+Footer: "[organizer / url]"
+
+A single bold geometric graphic element anchors the composition. Crisp kerning, precise alignment, print-grade. Avoid clip-art, gradients, drop shadows, fake sponsor logos.
+```
+
+### 3 · 移动 App UI mockup
+- 推荐:`portrait` (1024x1536, 含 1290x2796 画布感) · `high`
+- 示范:读起来像产品规格 · 虚构产品名避真品牌 · 顶栏含精确 in-image 文案
+
+```text
+Design a production-quality mobile app screen mockup, 1290x2796 smartphone canvas, modern fintech aesthetic, dark UI (#0d1117 base, #161b22 cards), one electric-blue accent #2f81f7.
+
+Top header with exact in-image text: "[App名]", "[问候语]", balance "[¥金额]".
+Below: a row of 4 quick-action chips "[A]" / "[B]" / "[C]" / "[D]"; an asset donut chart with legend; a 7-day line chart; a transactions list with [3-4] rows showing merchant, time and signed amount (gains gold, losses coral).
+Crisp typography, precise icon alignment, consistent 8pt spacing, realistic but invented data. No real brand logos, no garbled text.
+```
+
+### 4 · Web SaaS 仪表盘
+- 推荐:`landscape` (1536x1024) · `high`
+- 示范:16:10 画布 · 组件系统 · 可信数据 · publication-grade 清洁度
+
+```text
+Design a 16:10 desktop SaaS dashboard UI, clean light theme with a calm slate/indigo palette, left sidebar nav + top bar, card-based content grid.
+Exact in-image text: product name "[名]" top-left; nav items "[概览]" / "[分析]" / "[用户]" / "[设置]"; 4 KPI cards with label + big number + small delta (e.g. "活跃用户" 12,480 ▲3.2%); one large time-series area chart titled "[标题]"; a data table with [4] columns and [4] rows of plausible values.
+Crisp typography, precise alignment, generous whitespace, production-quality. Avoid clutter, fake logos, unreadable microtext.
+```
+
+### 5 · 科普信息图
+- 推荐:`portrait` 或 `landscape` · `high`
+- 示范:命名制品类型 · 固定布局区域 · 注释行为 · 风格边界
+
+```text
+Create a museum-catalog-style educational infographic about [主题]. Editorial science-poster look — muted academic palette, clean line work, NOT a generic ad poster.
+Fixed layout regions: top title band with "[标题]" and "[副标题]"; left column a labeled diagram of [对象] with numbered callout lines; right column 3 summary cards "[要点1]" / "[要点2]" / "[要点3]"; bottom a legend.
+Thin leader lines connect labels to parts; small magnified detail insets where useful. All labels exact, large and legible. Calm, uncluttered, classroom-wall clarity. Avoid photorealism, neon, garbled text.
+```
+
+### 6 · 学术 / 系统架构图
+- 推荐:`landscape` (1536x1024) · `high`
+- 示范:方向+会场风格 · 结构基元+有向关系语义 · publication-grade
+
+```text
+Landscape 16:9 academic system-architecture figure, NeurIPS camera-ready style — white background, clean boxes, large readable labels, uncluttered.
+Structure: [N] horizontal layers/columns labeled "[L1]", "[L2]", "[L3]"; inside each, named component boxes "[组件…]". Directed arrows show data flow left-to-right; a dashed feedback arrow returns from "[X]" to "[Y]"; a side panel labeled "[内存/注册表]". Color encodes role (one hue per layer), arrow style encodes relation (solid=data, dashed=control).
+Publication-grade, precise alignment, no decorative gradients. Avoid clutter and tiny unreadable text.
+```
+
+### 7 · 产品渲染(JSON 配置式)
+- 推荐:`portrait` 或 `square` · `high`
+- 示范:结构化 JSON 拆视觉子系统 · render flags · 高级电商质感
+
+```text
+/* PRODUCT_RENDER_CONFIG: [产品短名]  VERSION: 1.0  AESTHETIC: Premium Commercial Photography */
+{
+  "GLOBAL": { "aspect": "3:4", "render_flags": ["8K_UHD", "sharp_foreground", "micro_texture", "no_CGI_tell"] },
+  "ENVIRONMENT": { "surface": "[材质,如 wet slate]", "background": "[soft gradient studio]", "props": ["[道具1]", "[道具2]"] },
+  "CORE_ASSET": { "object": "[产品]", "materials": ["[材质细节]"], "label": "tiny generic mark only, no real brand" },
+  "LIGHTING": { "key": "softbox 45°", "rim": "cool blue", "accent": "warm copper highlight" },
+  "DYNAMICS": { "particles": "[如 condensation droplets / flour dust]" },
+  "OUTPUT": { "finish": "editorial", "palette": "[muted teal / bone / brass]" }
+}
+Render as a single hero product shot, low three-quarter angle. No visible real brand logos, no garbled labels.
+```
+
+### 8 · 风格化插画
+- 推荐:`square` 或 `portrait` · `medium`→`high`
+- 示范:风格锚点具体到媒介/制作语境 · 原创角色 · 线/着色方向
+
+```text
+A [画幅] stylized illustration in a [具体媒介风格,如 modern flat 2D with grainy risograph texture / painterly semi-realism] aesthetic.
+Subject: [原创角色或场景的具体描述,含动作与情绪]. Setting: [环境,5-8 个具体名词]. 
+Palette: [3 色,如 indigo / coral / cream]. Lighting: [方向与气质]. Composition: [构图,如 rule-of-thirds, character lower-left, negative space upper-right].
+Clean confident linework, [cel-shading / soft gradient] coloring. Original character — avoid copying any existing IP. Avoid AI-glossy plastic look, avoid muddy colors.
+```
+
+### 9 · 写实摄影
+- 推荐:`landscape` 或 `portrait` · `high`
+- 示范:单一主导捕获语境 · 场景密度(5-12 名词)胜形容词 · 反 AI 味
+
+```text
+A candid documentary-style photo of [场景,如 a convenience store at night], shot RAW, unprocessed, natural full-frame look, [一个主导机位,如 eye level from across the street with a 35mm feel].
+Scene density (specific nouns): [列 8-12 个,如 freezer condensation, flickering tube light, a hooded customer, wet asphalt reflections, a delivery e-bike, a vending machine glow, scattered receipts]. 
+Mixed lighting: [冷荧光 vs 暖钠灯]. Ordinary imperfect realism — natural skin, real-world wear, no posing.
+Avoid: cinematic color grade, studio light, HDR over-processing, plastic AI skin, staged composition.
+```
+
+### 10 · 角色设计参考单
+- 推荐:`landscape` (1536x1024) · `high`
+- 示范:多面板一致性 · 前/侧/后视图 · 共享美术指导
+
+```text
+An official-style character reference sheet for an original character, laid out as a clean grid on a neutral studio background.
+Panels: front / side / back full-body turnaround (consistent proportions and outfit across all three); a row of 4 facial expression close-ups "[喜]" "[怒]" "[惊]" "[平静]"; one prop/detail breakout; a small color palette swatch strip.
+Shared art direction: [风格锚点], consistent costume pattern, consistent lighting and line weight across panels. Character identity must stay identical in every panel. Original design — avoid resembling existing IP. Clean, model-sheet clarity.
+```
+
+### 11 · 等距 3D 微缩场景
+- 推荐:`square` (1024x1024) · `high`
+- 示范:等距投影约束 · 微缩可玩细节 · 柔光黏土质感
+
+```text
+A cute isometric 3D miniature diorama of [场景,如 a cozy developer's home office], true isometric projection, soft clay-render look, gentle global illumination.
+Packed with specific tasteful details: [列 8-10,如 a mechanical keyboard, dual monitors with tiny code, a coffee mug, a potted plant, bookshelves, warm desk lamp, a sleeping cat]. 
+Palette: [warm pastel, 3-4 色]. Soft shadows, subtle ambient occlusion, miniature tilt-shift feel. Clean background, centered composition. Avoid harsh contrast, avoid clutter beyond the diorama.
+```
+
+### 12 · Logo / 品牌系统板
+- 推荐:`square` 或 `landscape` · `high`
+- 示范:一板含 logo+字标+色板+应用 · 原创无真品牌
+
+```text
+A clean brand identity presentation board on an off-white background, for an invented brand "[名]" in the [行业] space.
+Show, in a tidy grid: the primary logomark (a simple original geometric symbol), the wordmark "[名]", a 4-swatch color palette with hex codes, a type-pairing sample (heading + body), and 2 small application mockups (e.g. a business card and an app icon).
+Minimal, modern, confident; consistent grid and spacing. Original mark only — no resemblance to real-world brands, no stock-clip-art. Crisp, print-grade.
+```
+
+---
+
+## 怎么用这个库(给 agent)
+
+1. 按需求品类在路由索引找最近的 1-2 条。
+2. 把方括号 `[…]` 占位换成本次需求的具体内容,保留骨架与法则。
+3. 叠加个人层:用 `profile` 的风格锚点/否定项覆盖默认,用 `samples search` 的采纳样本
+   微调措辞。**baseline 已好,个人层只调 delta。**
+4. 给出推荐 `size`·`quality`。

package/references/golden-cases.jsonl

@@ -0,0 +1,5 @@
+{"id":"poster-tea-zh","request":"茶饮新品海报，国风一点，写冷泡系列，价格 16/19 元，画面高级克制","asset_type":"poster","text":["冷泡系列","16/19 元"],"strict_text":true,"out":"tea-poster.png"}
+{"id":"ui-ai-gallery","request":"手机 App 首页 UI，产品叫 AI 画廊，展示热门作品、生成按钮、最近项目，风格极简","asset_type":"ui","text":["AI 画廊"],"strict_text":true,"out":"ai-gallery-ui.png"}
+{"id":"diagram-rag","request":"生成一张 RAG 系统架构图，白底论文风","asset_type":"diagram","subject":"RAG system architecture with five modules","text":["User","Retriever","Vector DB","LLM","Answer"],"strict_text":true,"out":"rag-diagram.png"}
+{"id":"product-smart-cup","request":"画一个智能水杯的电商产品渲染图，黑色磨砂材质，带温度显示，适合新品首发","asset_type":"product","out":"smart-cup.png"}
+{"id":"infographic-sleep","request":"做一张科普信息图，主题是睡眠周期，包含浅睡、深睡、REM 三个模块","asset_type":"infographic","text":["浅睡","深睡","REM"],"strict_text":true,"out":"sleep-infographic.png"}

package/references/harness.md

@@ -0,0 +1,134 @@
+# harness.md —— 三层偏好学习机制（越用越懂你）
+
+本 skill 的"自进化"靠三层数据。它们都在**数据目录**（不进 git repo）：
+
+```
+默认 ~/.local/share/draw-prompt/        # 可用环境变量 DRAW_PROMPT_HOME 覆盖
+├── style_profile.md      # ① 偏好档案：长期口味
+├── samples.jsonl         # ② 样本库：每次出图的采纳/弃用（few-shot 来源）
+└── judgements.jsonl      # ③ 评分回路：成品图质量轨迹（可选）
+```
+
+核心闭环：**每次编译前读三层 → 编译 → 出图（下游）→ 反馈写回三层 → 下次更准**。
+
+---
+
+## ① 偏好档案 style_profile.md
+
+人可读的 markdown：yaml frontmatter（结构化字段）+ body（自由笔记）。
+
+```markdown
+---
+default_aspect: "3:4"           # 没指定画幅时的默认
+default_quality: high           # 输出建议的默认 quality
+favored_styles: ["New Chinese", "cinematic"]   # 常用风格锚点
+avoided_elements: ["高饱和渐变", "假 logo"]      # 永远进否定项
+text_language: zh               # 文字默认语种
+updated_at: 2026-05-30T...Z
+---
+
+# 风格偏好笔记
+- (2026-05-30) 偏好大留白、克制配色，排斥高饱和渐变。
+- (2026-05-30) 海报类喜欢把主标题压到画面下三分之一。
+```
+
+**读**：每次编译第一步 `profile show`。
+**写**：
+- 结构化字段：`profile set favored_styles "A, B, C"`（列表字段逗号分隔）
+- 自由观察：`profile note "一句带依据的口味观察"`
+- 拿路径自己编辑：`profile path` → 直接改文件也行
+
+**写入纪律**：只有当一条偏好**反复出现、足够稳定**时才沉淀到档案。一次性的临时要求
+不要写进来，避免污染长期口味。每条笔记尽量带一句依据和日期。
+
+---
+
+## ② 样本库 samples.jsonl
+
+每行一条 JSON，记录一次出图的完整上下文：
+
+```json
+{"id":"20260530-053451-a1b2","ts":"...","request":"茶饮新品海报",
+ "prompt":"Design a 3:4 vertical poster...","verdict":"accept","reason":"",
+ "image":"out.png","size":"portrait","quality":"high","tags":["poster","chinese","tea"]}
+```
+
+`size`/`quality` 记录这次出图用的画幅与质量档（`samples add --size --quality` 写入），
+复盘时不丢失第七段的输出决策。
+
+- `verdict`：`accept` / `reject` / `pending`（出图后待反馈）
+- **写**：`samples add --request "…" --prompt "…" --verdict accept --tags "…"`
+- **反馈**：`feedback last --verdict reject --reason "构图太满"`（更新最近一条或指定 id）
+- **检索（few-shot）**：`samples search "poster chinese"`——采纳样本自动加权优先，
+  按关键词/标签匹配 request+prompt+tags。
+
+**怎么帮编译**：第 3 步检索同类采纳样本，借用其被验证有效的结构、措辞、否定项；
+reject 样本是负例，提醒这次规避同样的坑。
+
+**tags 约定**：用稳定的小写英文标签（资产类型 + 风格 + 主题），如
+`poster,chinese,tea` / `ui,fintech,dark` / `infographic,anatomy`，方便检索聚类。
+
+---
+
+## ③ 评分回路 judgements.jsonl（默认兜底）
+
+**默认每次出图都跑**：只要成品图在本地可读，agent 就用视觉按 rubric 自评一遍，作为被动
+学习的质量门——明显翻车的自动不采纳，不依赖用户开口。**评分由 agent（具备视觉）做，
+CLI 只存储**——
+本 skill 不看图、不调用 Codex、不引入任何额外 API。
+
+四轴 rubric（`judge rubric` 可查）：
+
+| 轴 | 含义 | 取值 |
+|---|---|---|
+| `text_accuracy` | 图中文字与 prompt 是否逐字一致、无乱码（无文字给 10） | 0-10 |
+| `composition` | 构图/布局/层级是否清晰 | 0-10 |
+| `style_match` | 是否贴合 prompt 的风格锚点 | 0-10 |
+| `artifacts` | 反向分，10=无瑕疵，0=明显畸变/多指/糊 | 0-10 |
+| `verdict` | 任一轴 ≤5 则 `revise`，否则 `pass` | pass/revise |
+| `failure_category` | revise 时的结构化失败分类，用于 `revise --reason` | text_error/layout_error/... |
+
+流程：使用者把成品图发回 → 你看图对照 prompt 打分 → 写入：
+```
+judge record --image <path> --prompt "<prompt>" \
+  --score '{"text_accuracy":9,"composition":8,"style_match":9,"artifacts":9,"verdict":"pass","notes":"一句话"}'
+```
+`verdict=revise` 时，主动指出是哪一轴拖后腿，填 `failure_category`，然后用
+`revise --sample-id <id|last> --reason <failure_category>` 生成下一版。
+
+评分轨迹（`judge list`）能看出某类需求的长期质量走向，是后续接 cron 自动巡检的地基
+（当前默认手动，不自动跑）。
+
+---
+
+## 被动学习：零打扰地获取偏好（核心原则）
+
+显式 accept/reject 摩擦太大——用户出完图通常不会回来评价，主动追问"满意吗"更是打扰。
+**所以学习默认从交互中被动发生，agent 不应追问满意度。** 四个零打扰信号源：
+
+| 信号 | agent 的动作 | 打扰用户? |
+|---|---|---|
+| 拿到成品图（`--out` 路径在本地可读） | **视觉质检兜底（默认必跑）**：agent 读图按 rubric 自评，明显翻车（文字乱码/畸变/糊）自动判 `reject`、不入 few-shot——不等用户开口 | 否 |
+| 出图后用户没表达不满，且质检通过 | 样本记 `accept`；但风格仍 `inferred`，**沉默不固化为默认口味** | 否 |
+| 用户提需求/追加要求时的措辞、否定 | 即时 `profile note`/`profile set` | 否 |
+| 同一需求多轮迭代、换版 | 被替代版自动 `feedback … --verdict reject`（原因取自用户的话），终版按质检结果 | 否 |
+
+规则：
+- **质检兜底、沉默不固化**。采纳前先过 agent 视觉质检这道门：质检通过且用户没反对 →
+  记 `accept`；明显翻车 → 自动弃，不问用户；用户主动表达不满 → 翻 `reject` 记原因。
+  但"采纳"只表示"这版质量过关、可作 few-shot"，**不等于把它的风格定成你的默认口味**——
+  风格要等复现或你明确认可才固化。用户零额外动作。
+- **图读不到则降级**：若成品图不在本地（用户没回传、路径不可达），质检跳过，样本暂存
+  `pending` 不固化，等下次有图或用户表态。绝不因为缺图就乱猜。
+- **标注来源与置信度**。写 profile 时注明 `observed`（用户明确说过的事实）还是
+  `inferred`（agent 从一次选择推断的假设）。**inferred 的偏好不要立刻当默认**——
+  等它在后续需求里复现、或没有被否定，再固化为 `favored_styles`/`default_*`。
+  这样避免把"agent 自己选的方向"误当成"用户的口味"。
+- **不要为了学习而打断创作**。沉淀是后台动作，对话里只在用户想纠正时才围绕偏好展开。
+
+## 渐进式启用
+
+- **默认**：①+②+③ 全开——偏好档案 + 样本库 + 每次出图视觉质检兜底。质检过滤翻车版，
+  沉默不固化口味，用户零打扰。
+- **未来**：评分轨迹 + 版本化 prompt 可挂定时巡检，自动提示哪类 prompt 该升级——
+  本 skill 已把数据格式备好，留作演进。

package/references/visual-cases.jsonl

@@ -0,0 +1,8 @@
+{"id":"poster-strict-text","scenario":"poster_text_overlay","request":"茶饮新品海报，透明冷泡茶瓶、山间晨雾，写冷泡系列，价格 16/19 元","asset_type":"poster","text":["冷泡系列","16/19 元"],"strict_text":true}
+{"id":"long-input-diagram","scenario":"long_input_to_diagram","request":"长输入转化 skill 的系统架构图，包含 User、Prompt Compiler、Template Router、Visual Plan、Handoff","asset_type":"diagram","text":["User","Prompt Compiler","Template Router","Visual Plan","Handoff"],"strict_text":true}
+{"id":"series-product-hero","scenario":"series_consistency_product","request":"茶饮品牌春季主视觉，透明冷泡茶瓶、山茶花、晨雾，保持高级统一视觉语言","asset_type":"product","style":"clean premium beverage campaign, editorial product photography","palette":"moss green,ivory,clear glass"}
+{"id":"brand-kv","scenario":"brand_consistency_kv","request":"新品冷泡茶品牌主视觉，透明茶瓶、山间晨雾、山茶花，品牌名山川茶事","asset_type":"poster","text":["山川茶事"],"strict_text":true,"style":"original, consistent, premium but restrained brand system","palette":"moss green,ivory,clear glass"}
+{"id":"character-reference","scenario":"character_consistency_reference","request":"原创角色 Mira 设定表，短银发、琥珀色眼睛、青绿色探险外套、黄铜指南针吊坠，三视图和表情板","asset_type":"character","text":["Mira"]}
+{"id":"data-viz","scenario":"data_visualization","request":"展示四周转化率趋势，W1 12%, W2 18%, W3 21%, W4 28%，标题转化率趋势","asset_type":"infographic","text":["转化率趋势"],"strict_text":true}
+{"id":"multi-aspect-square","scenario":"multi_size_square","request":"茶饮新品海报，透明冷泡茶瓶、山间晨雾，写冷泡系列","asset_type":"product","aspect":"1:1","text":["冷泡系列"],"strict_text":true}
+{"id":"rewrite-safety","scenario":"risky_reference_rewrite","request":"做一个迪士尼风格的角色海报，好看高级","asset_type":"poster"}